Hamiltonian Dynamical Systems and Applications (NATO Science for Peace and Security Series B: Physics and Biophysics) 9781402069635, 1402069634

Citation preview

Hamiltonian Dynamical Systems and Applications

NATO Science for Peace and Security Series This Series presents the results of scientific meetings supported under the NATO Programme: Science for Peace and Security (SPS). The NATO SPS Programme supports meetings in the following Key Priority areas: (1) Defence Against Terrorism; (2) Countering other Threats to Security and (3) NATO, Partner and Mediterranean Dialogue Country Priorities. The types of meeting supported are generally "Advanced Study Institutes" and "Advanced Research Workshops". The NATO SPS Series collects together the results of these meetings. The meetings are coorganized by scientists from NATO countries and scientists from NATO's "Partner" or "Mediterranean Dialogue" countries. The observations and recommendations made at the meetings, as well as the contents of the volumes in the Series, reflect those of participants and contributors only; they should not necessarily be regarded as reflecting NATO views or policy. Advanced Study Institutes (ASI) are high-level tutorial courses intended to convey the latest developments in a subject to an advanced-level audience Advanced Research Workshops (ARW) are expert meetings where an intense but informal exchange of views at the frontiers of a subject aims at identifying directions for future action Following a transformation of the programme in 2006 the Series has been re-named and re-organised. Recent volumes on topics not related to security, which result from meetings supported under the programme earlier, may be found in the NATO Science Series. The Series is published by IOS Press, Amsterdam, and Springer, Dordrecht, in conjunction with the NATO Public Diplomacy Division. Sub-Series A. B. C. D. E.

Chemistry and Biology Physics and Biophysics Environmental Security Information and Communication Security Human and Societal Dynamics

http://www.nato.int/science http://www.springer.com http://www.iospress.nl

Series B: Physics and Biophysics

Springer Springer Springer IOS Press IOS Press

Hamiltonian Dynamical Systems and Applications

edited by

Walter Craig McMaster University, Hamilton, ON, Canada

Published in cooperation with NATO Public Diplomacy Division

Proceedings of the NATO Advanced Study Institute on Hamiltonian Dynamical Systems and Applications Montreal, Canada 18–29 June 2007

Library of Congress Control Number: 2008920287

ISBN 978-1-4020-6963- 5 (PB) ISBN 978-1-4020-6962-8 (HB) ISBN 978-1-4020-6964 -2 (e-book)

Published by Springer, P.O. Box 17, 3300 AA Dordrecht, The Netherlands.

www.springer.com

Printed on acid-free paper

All Rights Reserved © 2008 Springer Science + Business Media B.V. No part of this work may be reproduced, stored in a retrieval system, or transmitted in any form or by any means, electronic, mechanical, photocopying, microfilming, recording or otherwise, without written permission from the Publisher, with the exception of any material supplied specifically for the purpose of being entered and and executed on a computer system, for exclusive use by the purchaser of the work.

Preface

This volume is a collection of lecture notes from the courses that were given during the 2007 Séminaire de Mathématiques Supérieure in Montréal (SMS), which was conceived and supported as a NATO Advanced Study Institute. The courses took place during the two-week period from June 18 to June 29, 2007, at the Centre de Recherches Mathématiques (CRM), and they were funded by a grant from NATO and from the ISM, which is the combined graduate mathematics program of the Montréal area. The organising committee for this event was D. Bambusi (Milan), W. Craig (McMaster), S. Kuksin (Edinburgh and Paris), and A. Neishtadt (Moscow). There were more than 80 participants, coming from around the world, and in particular there were a good number of students from France, from Italy, from Spain, from the United States and from Canada. The program of lectures occupied two complete weeks, with five or six one-hour lectures each day, so that in total 57 h of courses were presented. The topic of the 2007 NATO-ASI was Hamiltonian dynamical systems and their applications, which concerns mathematical problems coming from physical and mechanical systems of evolution equations. Many aspects of the modern theory of the subject were covered; topics of the principal lectures included low dimensional problems as well as the theory of Hamiltonian systems in infinite dimensional phase space, and and their applications to problems in classical mechanics, continuum mechanics, and partial differential equations. Applications were also presented to several important areas of research, including to celestial mechanics, control theory, the partial differential equations of fluid dynamics, and the theory of adiabatic invariants. It is a good thing to do to articulate the relevance of the subject matter of these SMS lectures to the physical sciences. Physical laws are for the most part expressed in terms of differential equations, and the most natural classes of these are in the form of conservation laws or of problems of the calculus of variations for an action functional. These problems can often be posed as Hamiltonian systems, whether dynamical systems on finite dimensional phase space as in classical mechanics, or partial differential equations (PDE) which are naturally of infinitely many degrees of freedom. For instance, the well known N-body problem of celestial mechanics is v

vi

Preface

still of great relevance to modern mathematics and more broadly to science; indeed in applications the mission design of interplanetary exploration regularly uses the gravitational boost of close encounters to manoeuvre their spacecraft (first used in the Mariner–10 mission, 1974). This is also true on the level of theoretical results, which can be traced to the work of Laplace, Lagrange and Poincaré, but whose modern successes date to the celebrated theory of Kolmogorov, Arnold and Moser (KAM) (1954/1961/1963). Recent mathematical progress includes the discoveries of new choreographies of many body orbits (Chenciner & Montgomery, 2000), and the constructions of Poincaré’s second species orbits (Bolotin & MacKay, 2001). Furthermore, the development of rigorous averaging methods (Nekhoroshev 1979) gives hope for realistic long time stability results (Neishtadt 1981, Treschev 1996, Pöschel 1999). Additionally, the last several years has seen major progress in the long outstanding problem of Arnold diffusion, with the advent of Mather’s variational techniques (2003) related to a generalised Morse–Hedlund theory, including Cheng’s subsequent work on variational methods, and the geometrical approach to the ‘gap problem’ due to de la Llave, Delshams & Seara (2006). Over the last decade the field of Hamiltonian systems has taken on completely new directions in the extension of the analytical methods of Hamiltonian mechanics to partial differential equations. The results of Kuksin, Wayne, Pöschel, Craig, Bambusi and Bourgain have introduced a new paradigm to the study of partial differential equations of evolution, where research focuses on the fundamental structures invariant under the dynamics of the PDE in an appropriate phase space of functions. Two basic examples of this direction of enquiry include (i) the development of several approaches to a KAM theory, with very recent contributions by Yuan (2006) and Eliasson & Kuksin (2007), and (ii) Nekhoroshev stability results for systems with infinitely many degrees of freedom (Bambusi 1999). These considerations show an exciting and extremely promising connection between Hamiltonian dynamical systems and harmonic analysis techniques in PDE. A case in point is the relationship between upper bounds on the growth of higher Sobolev norms of solutions of nonlinear evolution equations, and the bounds on orbits given by Nekhoroshev theory; similarly there is a possibly surprising connection between lower bounds on such growth and the existence of solution of PDE which exhibit phenomena related to Arnold diffusion. This research area of evolution equations and Hamiltonian systems is one of the most active and exciting fields of PDE in the last several years. The subjects in question involve by necessity some of the most technical aspects of analysis coming from a number of diverse fields, and before our event there has not been one venue nor one course of study in which advanced students or otherwise interested researchers can obtain an overview and sufficient background to enter the field. What we have done with the Montréal Advanced Studies Institute 2007 is to offer a series of lectures encompassing this wide spectrum of topics in PDE and dynamical systems. Most of the major developers in this field were speakers at this ASI, including the top international leaders in the subject. This has made it a unique opportunity for junior mathematicians to hear a focused set of lectures given by major researchers and contributors to the field. The organizers are grateful for the time and energy that the speakers devoted to the thoughtful preparation of

Preface

vii

their lectures, and to the subsequent written and complete versions that appear in this volume. And in addition the students at this ASI, who were for the most part advanced graduate students and postdoctoral fellows, included many very promising and active young mathematicians in the field, with their own well-developed research programs. The participants’ enthusiasm for the ASI, their help in writing lecture notes for the courses, and their general cheerfulness and good attitude during the course of the two weeks of lectures, made the event an experience not to be forgotten. Last but not least, the organizers of the SMS 2007 would like to acknowledge the generous and timely support of the Public Diplomacy Division of NATO, without which the two weeks of this Advanced Study Institute would not have taken place, the additional financial support of the Montréal Centre de Recherches Mathématiques (CRM), the ISM and the Université de Montréal, and for the dependable guidance and initiative of Sakina Benhima, our Directrice de Programme at the CRM in Montréal. The series of lectures in this volume includes the following topics: Hamiltonian systems and optimal control (A. Agrachev, SISSA, Trieste), Birkhoff normal form for some semilinear PDEs (D. Bambusi, Universita degli Studi di Milano), Variational methods for Hamiltonian PDEs (M. Berti, Università degli Studi di Napoli), The N-body problem (A. Chenciner, Observatoire de Paris), Variational methods for the problem of Arnold diffusion (C.-Q. Cheng, Nanjing University), The transformation theory of Hamiltonian PDE and the problem of water waves (W. Craig, McMaster University), Geometric approaches to diffusion and instability (R. de la Llave, University of Texas at Austin), KAM for the nonlinear Schrödinger equation (H. Eliasson, Université de Paris 7), Groups and topology in Euler hydrodynamics and the KdV (B. Khesin, University of Toronto), Three theorems on perturbed KdV (S. Kuksin, Heriot-Watt University), Averaging methods and adiabatic invariants (A. I. Neishtadt, Space Research Institute, Russian Academy of Science), Periodic KdV equation in weighted Sobolev spaces (J. Pöschel, Universität Stuttgart), The forced pendulum as a model for dynamical behavior (P. Rabinowitz, University of Wisconsin), Normal forms of holomorphic dynamical systems (L. Stolovitch, Université Paul Sabatier), Some aspects of finite dimensional Hamiltonian systems (D. Treschev, Moscow State University), Infinite dimensional dynamical systems and the Navier–Stokes equations (C. E. Wayne, Boston University), and KAM theory with applications to nonlinear wave equations. (X. Yuan, Fudan University). Hamilton and Montréal, Canada

Walter Craig July 2007

Contents

Some aspects of finite-dimensional Hamiltonian dynamics . . . . . . . . . . . . . D.V. Treschev 1 Symplectic structure. Invariant form of the Hamiltonian equations 1.1 Hamiltonian equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 The Poisson bracket . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3 Liouville theorem on completely integrable systems . . . . 2 A pendulum with rapidly oscillating suspention point . . . . . . . . . . . 3 Anti-integrable limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 The standard map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Anti-integrable limit . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Proof of the Aubry theorem . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Some remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 Separatrix splitting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1 Poincaré’s observation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 The Poincaré integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Proof of Theorem 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4 Standard example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Four lectures on the N-body problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Alain Chenciner 1 The Poincaré–Birkhoff–Conley twist map of the annulus for the planar circular restricted three-body problem . . . . . . . . . . . . 1.1 The Kepler problem as an oscillator . . . . . . . . . . . . . . . . . . 1.2 The restricted problem in the lunar case . . . . . . . . . . . . . . . 1.3 Hill’s solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4 The annulus twist map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 The Arnold–Herman stability theorem for the spatial (1 + n)-body problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 The secular Hamiltonian . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Herman’s normal form theorem and how to use it . . . . . . .

1 1 1 3 4 5 8 8 9 11 12 13 13 14 15 18 19 21

21 21 22 24 26 29 30 33

ix

x

Contents

2.3 A stability theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Herman’s degeneracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Minimal action and Marchal’s theorem . . . . . . . . . . . . . . . . . . . . . . . 3.1 Central configurations and their homographic motions . . . 3.2 Variational characterizations of Lagrange’s equilateral solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Marchal’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4 Minimization under symmetry constraints . . . . . . . . . . . . . 4 Global continuation via minimization . . . . . . . . . . . . . . . . . . . . . . . . 4.1 Bifurcations from the Lagrange equilateral relative equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 From the equilateral triangle to the Eight . . . . . . . . . . . . . . 4.3 From the square to the Hip-Hop . . . . . . . . . . . . . . . . . . . . . 4.4 The avatars of the regular n-gon relative equilibrium: eights, chains and generalized Hip-Hops . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Averaging method and adiabatic invariants . . . . . . . . . . . . . . . . . . . . . . . . . . Anatoly Neishtadt 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Adiabatic invariance in one-frequency systems . . . . . . . . . . . . . . . . . 3 On adiabatic invariance in multi-frequency systems . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Transformation theory of Hamiltonian PDE and the problem of water waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Walter Craig 1 Hamiltonian systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Partial differential equations as Hamiltonian systems . . . . . . . . . . . . 3 The problem of water waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 The Dirichlet–Neumann operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 Perturbation theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 The calculus of transformations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Three theorems on perturbed KdV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sergei B. Kuksin 1 KdV equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 Integrability of (KdV) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Normal forms for (KdV) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 KAM-theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 Averaging: Hamiltonian perturbations . . . . . . . . . . . . . . . . . . . . . . . . 4 Averaging: case of non-Hamiltonian perturbations . . . . . . . . . . . . . . 4.1 Deterministic perturbations . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Random perturbations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

35 36 37 37 38 40 42 43 44 46 48 50 50 53 53 54 63 65 67 67 68 71 73 74 75 82 85 85 86 87 88 89 90 90 90 91

Contents

xi

Groups and topology in the Euler hydrodynamics and KdV . . . . . . . . . . . . 93 Boris Khesin 1 Euler equations and geodesics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 1.1 The Euler hydrodynamics equation . . . . . . . . . . . . . . . . . . . 93 1.2 Geodesics on Lie groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94 1.3 Geodesic description for various equations . . . . . . . . . . . . 95 2 Topology of steady flows . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 2.1 Arnold’s classification of steady fluid flows . . . . . . . . . . . . 95 2.2 Variational principles for steady flows . . . . . . . . . . . . . . . . 97 3 Euler equations and integrable systems . . . . . . . . . . . . . . . . . . . . . . . 98 3.1 Hamiltonian reformulation of the Euler equations . . . . . . . 98 3.2 The Virasoro algebra and the KdV equation . . . . . . . . . . . 99 3.3 Equations–relatives and conservation laws . . . . . . . . . . . . . 100 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 Infinite dimensional dynamical systems and the Navier–Stokes equation . 103 C. Eugene Wayne 1 First lecture: infinite dimensional dynamical systems . . . . . . . . . . . 103 2 Second lecture: invariant manifolds for partial differential equations on unbounded domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 3 Third lecture: an introduction to the Navier–Stokes equations . . . . 122 4 Fourth lecture: the long-time asymptotics of solutions of the two-dimensional Navier–Stokes equation . . . . . . . . . . . . . . . . 129 5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140 Hamiltonian systems and optimal control . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 Andrei Agrachev 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143 2 First lecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 144 2.1 First order conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145 3 Second lecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 146 3.1 Second variation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 4 Third lecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 149 4.1 Curves in the Lagrange Grassmannians . . . . . . . . . . . . . . . 150 4.2 Curvature-type invariants . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 5 Fourth lecture . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 156 KAM theory with applications to Hamiltonian partial differential equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157 Xiaoping Yuan 1 Brief history and basic ideas of KAM theory . . . . . . . . . . . . . . . . . . 157 2 Derivation of the linearized equations . . . . . . . . . . . . . . . . . . . . . . . . 163 3 Solutions of the linearized equations . . . . . . . . . . . . . . . . . . . . . . . . . 167 4 Applications to partial differential equations in higher dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175

xii

Contents

Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176 Four lectures on KAM for the non-linear Schrödinger equation . . . . . . . . . 179 L.H. Eliasson and S.B. Kuksin 1 The non-linear Schrödinger equation . . . . . . . . . . . . . . . . . . . . . . . . 179 1.1 The non-linear Schrödinger equation . . . . . . . . . . . . . . . . . 179 1.2 An ∞ -dimensional Hamiltonian system . . . . . . . . . . . . . . . 180 1.3 The topology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182 1.4 Action-angle variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183 1.5 Statement of the result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184 1.6 KAM-tori . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 1.7 Consequences of Theorem 1 . . . . . . . . . . . . . . . . . . . . . . . . 186 1.8 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188 1.9 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 189 2 The homological equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190 2.1 Normal form Hamiltonians . . . . . . . . . . . . . . . . . . . . . . . . . 190 2.2 The KAM-iteration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 2.3 The components of the homological equation . . . . . . . . . . 192 2.4 Small divisors and the second Melnikov condition . . . . . . 194 3 Normal form Hamiltonians . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 3.1 Blocks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195 3.2 Lipschitz domains . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 197 3.3 Töplitz at ∞(d = 2) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199 3.4 Töplitz–Lipschitz matrices (d = 2) . . . . . . . . . . . . . . . . . . . 200 3.5 Normal form Hamiltonians . . . . . . . . . . . . . . . . . . . . . . . . . 201 4 Estimates of small divisors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202 4.1 A basic estimate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202 4.2 The second Melnikov condition (d = 2) . . . . . . . . . . . . . . . 203 5 Functions with the Töplitz–Lipshitz property (d = 2) . . . . . . . . . . . 208 5.1 Töplitz structure of the Hessian . . . . . . . . . . . . . . . . . . . . . . 208 5.2 Töplitz–Lipschitz matrices L × L → gl(2, R) . . . . . . . . . 209 5.3 Functions with the Töplitz–Lipschitz property . . . . . . . . . 211 5.4 A short remark on the proof of Theorem 1.1 . . . . . . . . . . . 211 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212 A Birkhoff normal form theorem for some semilinear PDEs . . . . . . . . . . . . 213 D. Bambusi 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213 2 Birkhoff’s theorem in finite dimensions . . . . . . . . . . . . . . . . . . . . . . . 214 2.1 Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 214 2.2 Proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216 3 The case of PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221 3.1 Hamiltonian formulation of the wave equation . . . . . . . . . 221 3.2 Extension of Birkhoff’s theorem to PDEs: heuristic ideas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223

Contents

xiii

4

A Birkhoff normal form theorem for semilinear PDEs . . . . . . . . . . . 224 4.1 Maps with localized coefficients and their properties . . . . 224 4.2 Statement of the normal form theorem and its consequences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228 4.3 Application to the nonlinear wave equation . . . . . . . . . . . . 229 5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 6 Proofs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 6.1 Proof of the properties of functions with localized coefficients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231 6.2 Proof of the Birkhoff normal form Theorem 4.3 and of its dynamical consequences . . . . . . . . . . . . . . . . . . . 236 6.3 Proof of Proposition 4.2 on the verification of the property of localization of coefficients . . . . . . . . . . . 240 6.4 Proof of Theorem 4.4 on the nonresonance condition . . . . 244 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246

Normal form of holomorphic dynamical systems . . . . . . . . . . . . . . . . . . . . . 249 Laurent Stolovitch 1 Definitions and examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249 1.1 Vector fields and differential equations . . . . . . . . . . . . . . . . 250 1.2 Normal forms of vector fields . . . . . . . . . . . . . . . . . . . . . . . 253 1.3 Examples about linearization . . . . . . . . . . . . . . . . . . . . . . . . 256 1.4 Examples about nonlinearizable vector fields . . . . . . . . . . 258 2 Holomorphic normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259 2.1 Theorem of A.D. Brjuno . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259 2.2 Theorems of J. Vey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261 2.3 Singular complete integrability–Main result . . . . . . . . . . . 261 2.4 How to recover Brjuno’s and Vey’s theorems from Theorem 2.3.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 268 2.5 Sketch of the proof . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269 3 Proof of main Theorem 2.3.6 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273 3.1 Bounds for the cohomological equations . . . . . . . . . . . . . . 273 3.2 Iteration scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277 3.3 Proof of the theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278 4 Miscellaneous results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280 4.1 Normal forms again . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280 4.2 KAM theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281 4.3 Poisson structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 282 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283 Geometric approaches to the problem of instability in Hamiltonian systems. An informal presentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285 Amadeu Delshams, Marian Gidea, Rafael de la Llave, and Tere M. Seara 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285 1.1 Two types of geometric programs . . . . . . . . . . . . . . . . . . . . 287

xiv

Contents

2

Exposition of the Arnold example . . . . . . . . . . . . . . . . . . . . . . . . . . . 289 2.1 The obstruction property . . . . . . . . . . . . . . . . . . . . . . . . . . . 291 2.2 Some final remarks on the example in [Arn64] . . . . . . . . . 293 3 Return to a normally hyperbolic manifold. The two dynamics approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294 3.1 The basics of the mechanism of return to a normally hyperbolic invariant manifold . . . . . . . . . . . . . . . . . . . . . . . 294 3.2 The scattering map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296 3.3 The scattering map and homoclinic intersections of submanifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298 3.4 Monodromy of the scattering map . . . . . . . . . . . . . . . . . . . . 299 3.5 Smoothness and smooth dependence on parameters . . . . . 299 3.6 Geometric properties of the scattering map . . . . . . . . . . . . 300 3.7 Calculation of the scattering map . . . . . . . . . . . . . . . . . . . . 301 4 The large gap model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303 4.1 Generation of intersections. Melnikov theory for normally hyperbolic manifolds . . . . . . . . . . . . . . . . . . . 304 4.2 Computation of the scattering map . . . . . . . . . . . . . . . . . . . 307 4.3 The averaging method. Resonant averaging . . . . . . . . . . . . 308 4.4 Repeated averaging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312 4.5 Invariant objects generated by resonances: secondary tori, lower dimensional tori . . . . . . . . . . . . . . . . . . . . . . . . . 312 4.6 Heteroclinic intersections between the invariant objects generated by resonances . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315 5 The method of correctly aligned windows . . . . . . . . . . . . . . . . . . . . . 316 6 The large gap model: the method of correctly aligned windows . . . 318 7 The large gap model in higher dimensions . . . . . . . . . . . . . . . . . . . . . 319 8 Instability caused by normally hyperbolic laminations . . . . . . . . . . . 320 8.1 Models with two time scales: geodesic flows, billiards with moving boundaries, Littlewood problems . . . . . . . . . 321 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 324 Appendix . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330

Variational methods for the problem of Arnold diffusion . . . . . . . . . . . . . . 337 Chong-Qing Cheng 1 Introduction to Mather theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 337 2 Existence of Homoclinic Orbits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 341 3 Pseudo-connecting orbit set C˜η ,µ ,ψ . . . . . . . . . . . . . . . . . . . . . . . . . . 342 4 Existence of heteroclinic orbits . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345 5 Construction of global connecting orbits . . . . . . . . . . . . . . . . . . . . . . 353 6 Application to a priori unstable systems . . . . . . . . . . . . . . . . . . . . . . . 364 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 365

Contents

xv

The calculus of variations and the forced pendulum . . . . . . . . . . . . . . . . . . . 367 Paul H. Rabinowitz 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367 2 Periodic solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 369 3 Heteroclinic solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374 4 Multitransition solutions: the simplest case . . . . . . . . . . . . . . . . . . . . 380 5 Multitransition solutions: general case . . . . . . . . . . . . . . . . . . . . . . . . 387 6 The tip of the iceberg . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 390 Variational methods for Hamiltonian PDEs . . . . . . . . . . . . . . . . . . . . . . . . . . 391 Massimiliano Berti 1 Finite dimensions: resonant center theorems . . . . . . . . . . . . . . . . . . . 391 1.1 The variational Lyapunov–Schmidt reduction . . . . . . . . . . 393 2 Infinite dimensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 396 3 The variational Lyapunov–Schmidt reduction . . . . . . . . . . . . . . . . . . 397 3.1 The bifurcation equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 399 4 The small divisor problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 402 5 The (Q1)-equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 406 6 A variational principle on a Cantor set . . . . . . . . . . . . . . . . . . . . . . . . 408 7 Forced vibrations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412 7.1 The variational Lyapunov–Schmidt reduction . . . . . . . . . . 413 7.2 The bifurcation equation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 419 Spectral gaps of potentials in weighted Sobolev spaces . . . . . . . . . . . . . . . . . 421 Jürgen Pöschel 1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 421 2 Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 425 3 Gap Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426 4 Coefficient Estimates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 426 5 Modified Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 429 On the well-posedness of the periodic KdV equation in high regularity classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431 Thomas Kappeler and Jürgen Pöschel 1 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431 2 Birkhoff Coordinates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 435 3 Regularity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440

Some aspects of finite-dimensional Hamiltonian dynamics D.V. Treschev∗

Abstract These lectures touch upon two aspects of Hamiltonian mechanics. The first one (geometric) establishes fundamental role of symplectic geometry as the language of Hamiltonian mechanics. The second aspect (dynamical) exhibits the main problem in the domain, which is the interplay between regular and chaotic motion.

1 Symplectic structure. Invariant form of the Hamiltonian equations 1.1 Hamiltonian equations Hamiltonian system1 is an ODE-system which in certain coordinates q=(q1 , . . . , qn ), p = (p1 , . . . , pn ) can be presented in the form q˙ =

∂H , ∂p

p˙ = −

∂H , ∂q

(1)

the function H(q, p) is called the Hamiltonian function. Frequently, nonautonomous systems are considered, where H = H(q, p,t). This definition looks very non-geometrical. Although all calculations are anyway presented in coordinates (partially we will see this below), it would be good to present an equivalent invariant (coordinate independent) definition. Recall that a symplectic structure on a smooth manifold M is a closed nondegenerate differential two-form ω . The pair (M, ω ) is a symplectic manifold.

∗

Steklov Mathematical Institute e-mail: [email protected].

1

We will consider only the case of Hamiltonian ODE’s.

W. Craig (ed.), Hamiltonian Dynamical Systems and Applications, 1–19. c 2008 Springer Science + Business Media B.V.


1

2

D.V. Treschev

Theorem 1 (Darboux) In a neighborhood of any point of M there are local coordinates (q, p) = (q1 , . . . , qn , p1 , . . . , pn ), in which the symplectic structure has the form ω = d p ∧ dq. Corollary 1 Any symplectic manifold is even dimensional. Such coordinates (q, p) are called symplectic, canonical, or Darboux coordinates. Note that ω associates to any vector field v on M the differential 1-form f : f (·) = ω (·, v), where on the empty place · an arbitrary vector field can be posed. Let J be the inverse operator. It exists because ω is non-degenerate, and the dimensions of the vector spaces Tx M and Tx∗ M (x ∈ M) coincide. Then f (·) = ω (·, J f ). Let H : M → R be a smooth function. It determines the 1-form dH. Definition 1 The vector field vH = JdH on M is called the Hamiltonian vector field with Hamiltonian H. Hence dH(·) = ω (·, vH ). Problem 1 Check that in canonical coordinates the Hamiltonian vector field takes the traditional form vH = (Hp , −Hq ). Any map T : M → M preserving the symplectic structure is called symplectic. Symplectic maps can be regarded as discrete analogs of Hamiltonian systems. Problem 2 Let (q, p) be canonical local coordinates on M and let T : M → M be symplectic. Prove that the functions (P, Q) = (q ◦ T, p ◦ T ) are also canonical local coordinates on M. Problem 3 Let (q, p) and (P, Q) be local coordinates on M such that for some smooth function W = W (q, P) p=

∂W , ∂q

Q=

∂W . ∂P

(2)

Suppose also that the coordinates (q, p) are canonical. Prove that (P, Q) are also canonical. Hence in variables P, Q equations (1) remain the same:

∂H Q˙ = , ∂P

∂H P˙ = − , ∂Q

where the Hamiltonian is the same: H (P, Q) = H(p, q).

Some aspects of finite-dimensional Hamiltonian dynamics

3

The function W can depend on t. It is called a generating function of the canonical transformation (q, p) → (P, Q). In the non-autonomous case the new Hamiltonian equals H (P, Q,t) = ∂ W (q, P,t)/∂ t + H(p, q,t).

1.2 The Poisson bracket Let (M, ω ) be a symplectic manifold. For any two functions H, F on M we define the Poisson bracket {H, F} := ∂vH F = dF(vH ). Here ∂vH is the operator of differentiation w.r.t. the vector field vH . The first equality is a definition, while the second one is just an identity. We have the following simple properties of the Poisson bracket. 1. A smooth function F is a first integral of the Hamiltonian equations with Hamiltonian H ⇐⇒ {H, F} = 0. 2. {H, F} = ω (vH , vF ). 3. The operation {·, ·} is bilinear and skew-symmetric. According to 1 and 3 in any (autonomous) Hamiltonian system the Hamiltonian is a first integral.
 4. In canonical coordinates {H, F} = ∑nj=1 ∂∂ pHj ∂∂qFj − ∂∂ qHj ∂∂ pFj . A direct calculation in canonical coordinates gives 5. The Leibnitz identity: {FG, H} = F{G, H} + {F, H}G. 6. The Jacobi identity: {F, {G, H}} + {G, {H, F}} + {H, {F, G}} = 0 for any three functions F, G, H : M → R. This Poisson bracket is non-degenerate, i.e., for any z ∈ M and any function F such that dF = 0 at z there exists G such that {F, G} = 0. In some physical problems degenerate Poisson brackets appear,2 but we will not deal with these cases below. For any two vector fields u, v on M let [u, v] be their commutator:

∂[u,v] = ∂u ∂v − ∂v ∂u . Theorem 2 For any two functions F, G on M [vF , vG ] = v{F,G} . 2

Such Poisson brackets are not generated by symplectic structures.

4

D.V. Treschev

Proof. For an arbitrary function ϕ on M we have:

∂v{F,G} ϕ = {{F, G}, ϕ } = −{{G, ϕ }, F} − {{ϕ , F}, G} = {F, {G, ϕ }} − {G{F, ϕ }} = (∂vF ∂vG − ∂vG ∂vF )ϕ .



Proposition 1 (Poisson). Let F and G be first integrals of the Hamiltonian system (M, ω , H). Then {F, G} is also a first integral. Indeed, if {H, F} = {H, G} = 0 then by the Jacobi identity {H, {F, G}} = 0.

 Unfortunately, this statement is not of much use in the problem of the search for new integrals of motion. Usually the Poisson bracket of two integrals is an already known integral or zero. We say that two functions F, G are in involution or commute if {F, G} = 0.

1.3 Liouville theorem on completely integrable systems Suppose that the system (M, ω , H) (dim M = 2m) has m first integrals F1 , . . . , Fm in involution: {Fj , Fk } = 0. Consider the joint integral level M f = {z ∈ M : Fj (z) = f j = const, j = 1, . . . , m}.

(3)

Theorem 3 (Liouville–Arnold) Suppose that on M f the functions Fj are independent. Then 1. M f is a smooth manifold, invariant with respect to the Hamiltonian system z˙ = vH . 2. Each compact connected component of M f is diffeomorphic to an mdimensional torus 3 Tm . 3. In some coordinates (ϕ1 , . . . , ϕm ) mod 2π on Tm the Hamiltonian equations have the form ϕ˙ = λ , where λ = λ ( f ) ∈ Rm is a constant vector. Proof. Assertion (1) follows from the implicit function theorem. To check (2) and (3), we note that the vector fields v j = vFj are tangent to M f . (Indeed, ∂v j Fk = {Fj , Fk } = 0.) Since the functions Fj are independent on M f , the vector fields v j are also independent on M f . Moreover, [v j , vk ] = v{Fj ,Fk } = 0. It remains to use the following geometric fact (see for example, [2]): Lemma 2 Any compact connected m-dimensional manifold on which there are m everywhere independent commuting vector fields is diffeomorphic to the torus Tm . Moreover there are angular coordinates (ϕ1 , . . . , ϕm ) mod 2π on it such that all the 

m vector fields become constant (ν j = const ∈ Rm ). 3

In the non-compact case M f turns out to be Tk × Rm−k , 0
k < m (see [2]).

Some aspects of finite-dimensional Hamiltonian dynamics

5

Problem 4 Check that the tori Tmf from Theorem 3 are Lagrangian, i.e., dim Tmf = m and restriction of the symplectic structure to Tmf vanishes. Hamiltonian systems having a complete set (i.e., m) of almost everywhere independent first integrals in involution are said to be completely, or Liouville integrable. In Liouville integrable systems there are convenient, so-called, action-angle coordinates (ϕ , I) (I are the actions and ϕ are the angles) such that • ω = dI ∧ d ϕ (symplecticity), • H = H(I) (i.e., I are first integrals), • ϕ = ϕ mod 2π (i.e. ϕ are angular coordinates on the tori Mh ).

2 A pendulum with rapidly oscillating suspention point Mathematical pendulum is a (classical) mechanical system that consists of the rigid weightless rod AB with fixed end A. A point with mass m is attached to the end B. The motion is assumed to take place in a fixed vertical plane in the constant gravity force field. This system is well-known and Liouville integrable. Consider a more complicated problem. Let the point A vertically periodically oscillate. Period and amplitude of the oscillations is assumed to be small (of order ε ). We are interested in the action of the oscillations of the suspension point on the dynamics. Consider in the plane of motion a fixed coordinate system such that the x-axis is horizontal, the y-axis is vertical, and A lies on the y-axis. We assume that in this coordinate system 
ωt  g . , ω= A(t) = 0, aε cos ε l Here g is the gravity acceleration, l = |AB| is the length of the pendulum, and ε is small. The frequency ω is introduced so that ε is dimensionless. The dimension of a is length. The system is non-autonomous and has one degree of freedom. It is convenient to take the angle ϕ between the pendulum and the vertical, directed downward, as a variable, that determines position of the system. Problem 5 Obtain the Lagrangian of the system. Hint. L = T (ϕ , ϕ˙ ,t) − V (ϕ ,t), where T and V are kinetic and potential energy of the pendulum. Answer.
 ωt ωt  ωt m
2 2 l ϕ˙ − 2al ω ϕ˙ sin ϕ sin − mg aε cos + a2 ω 2 sin2 − l cos ϕ . L= 2 ε ε ε

6

D.V. Treschev

It is convenient to remove from L all terms which depend only on time and to divide L by ml 2 . Let Lˆ be the Lagrangian, obtained in this way:

ϕ˙ 2 aω ωt − ϕ˙ sin ϕ sin + ω 2 cos ϕ . Lˆ = 2 l ε Problem 6 Prove that Lagrangian systems with Lagrangians L and Lˆ are the same. Problem 7 Obtain the Hamiltonian of the system. Hint. H and Lˆ are related by the Legendre transform: H(ϕ , p,t) = pq˙ − ˆL(ϕ , ϕ˙ ,t), where ϕ˙ in the right-hand side should be expressed in terms of (ϕ , p,t) ˆ ∂ ϕ˙ . from the equation p = ∂ L/ Answer. ωt aω sin ϕ sin , p = ϕ˙ − l ε aω ω t a2 ω 2 2 ωt p2 +p sin ϕ sin + sin ϕ sin2 − ω 2 cos ϕ . H= 2 l ε 2l 2 ε We will construct a canonical change of variables which removes dependence of H on t in the main (zero) approximation in ε . We look for a change (ϕ , p) → (Φ , P) in the form
∂W ∂W ωt  , W = Pϕ + ε f ϕ , P, , , Φ= p= ∂ϕ ∂P ε where f is 2π -periodic in the last argument.4 We have: p = P + ε fϕ ,

Φ = ϕ + ε fP .

The new Hamiltonian reads


ωt  ωt  ωt  = ε ft + H ϕ , p, = ω D3 f + H Φ − ε fP , P + ε fϕ , , H Φ , P, ε ε ε where D3 is the derivative in the third argument. We obtain: H = ω D3 f (Φ , P, τ ) + H(Φ , P, τ ) + O(ε ),

τ=

ωt . ε

Therefore H does not depend on t in zero approximation in ε provided the function F = ω D3 f (Φ , P, τ ) + P

aω a2 ω 2 2 sin Φ sin τ + sin Φ sin2 τ l 2l 2

does not depend on τ . We choose a a2 ω f (Φ , P, τ ) = P sin Φ cos τ + 2 sin2 Φ sin 2τ , l 8l 4

This periodicity condition is necessary to have a change uniformly close to the identity for all t.

Some aspects of finite-dimensional Hamiltonian dynamics

and get5 : F =

a2 ω 2 4l 2

7

sin2 Φ . Hence, in the new variables H =

a2 ω 2 2 P2 − ω 2 cos Φ + sin Φ + O(ε ), 2 4l 2

where the part of H , contained in O(ε ), is 2π -periodic in τ . Remark 1 In fact it is possible to move the dependence on time to order O(ε N ) for an arbitrary N > 0, and even to O(e−c/|ε | ) for some positive constant c. However it is impossible to reach more: for any 2π -periodic in τ canonical near-identity change of variables the dependence of H on t will be greater than of order e−C/|ε | for a certain constant C > 0. Now let us study the system we have just obtained, neglecting the terms O(ε ). The system can be interpreted as the one describing the motion of a particle on a line (or on the circle Φ mod 2π ) in the force field with potential
 a2 V = ω 2 − cos Φ + 2 sin2 Φ . 4l The phase portrait of the system is (by definition) the set of level lines of the energy 2 integral P2 +V (Φ ) = const. As usual, it is convenient to draw it under the graph of the potential energy. There are two cases, see Fig. 1. The left-hand side of the figure contains the case of “small” amplitude a2 < 2l 2 . In this situation there are no qualitative differences with the case of the ordinary pendulum (a = 0).

Fig. 1 Phase portraits. Left: a2 < 2l 2 , and right: a2 > 2l 2

5

Recall once more that f should be periodic in τ .

8

D.V. Treschev

The situation changes drastically, when a2 > 2l 2 (the right-hand side of the figure). In this case a bifurcation occurs and the equilibrium Φ = ±π becomes stable. Moreover, the terms O(ε ) in the Hamiltonian do not destroy this effect, but we will not go into the detail. Problem 8 Draw the phase portrait in the case a2 = 2l 2 .

3 Anti-integrable limit 3.1 The standard map The standard map is, probably, the basic conceptual model for Hamiltonian dynamics in two degrees of freedom. Consider the cylinder Z = {(x, y) : x mod 2π } and its self-map Tε : Z → Z , (x, y) → Tε (x, y) = (X,Y ), where X = x + y + ε sin x,

Y = y + ε sin x.

(1)

Here ε is a real parameter which controls the type of the dynamics (regular or chaotic). The cylinder Z is said to be the phase space of the system. The dynamics should be understood as properties of the trajectories, i.e., sequences of points (xk , yk ) ∈ Z such that for any integer k (xk+1 , yk+1 ) = Tε (xk , yk ). The cylinder Z is a two-dimensional symplectic manifold with symplectic structure ω = dy ∧ dx. Problem 9 Check that the map Tε is symplectic, i.e., Tε∗ ω = ω . Any of you can easily look at trajectories of Tε by using a computer. To this end we remark that the variable y can be also regarded as angular. Indeed, Tε “respects” not only the shift of x by 2π , but also the analogous shift of y in the sense that for any integer k and n Tε (x + 2π k, y + 2π n) = (X + 2π k,Y + 2π k + 2π n) (shifts of X and Y also have the form 2π · (integer number)). Hence we can ask the computer to draw on the screen the square S = {(x, y) : 0
x
2π , 0
y
2π },

Some aspects of finite-dimensional Hamiltonian dynamics

9

to take an initial point (x0 , y0 ) ∈ S and to put it on the screen, to compute the point (x1 , y1 ) = Tε (x0 , y0 ) and to put it on the screen, etc. If a point (xn , yn ) leaves the square, it should be returned6 to S by the shift of x and/or y by 2π k with a proper integer k. I recommend you to do this and to look at the trajectories for various values of ε . Consider the case ε = 0. The system becomes a discrete analog of a Liouville integrable Hamiltonian system. The variables x, y play the role of action-angle variables. In particular, the action y is a first integral. Any trajectory lies on the curve (on the one-dimensional torus) lc = {(x, y) ∈ Z : y = c = const}. The curve lc rotates by the angle c. If c/π is rational, the trajectory is periodic. If c/π is irrational, the trajectory fills lc densely. Such curves lc are said to be non-resonant. In the case ε = 0 the situation gets much more complicated. One should not hope that any regular first integral exists, because trajectories (at least, some of them) stop to lie on smooth curves (like the circles lc ) and begin to demonstrate a chaotic behavior. However, the chaos appears gradually. According to the KAM-theory for small values of ε many of nonresonant curves lc , slightly deformed, exist as invariant curves for Tε . These curves can be easily seen on pictures, produced by numerical simulations. Trajectories, lying on these circles, are regarded as regular. Chaotic trajectories are presented on a computer screen as clouds, more or less densely filled with points. If ε is small and initial conditions are taken randomly, regular trajectories are more probable. When ε increases, the curves lc,ε destroy and chaos becomes more noticeable. For large ε numerical simulations show that a "typical" trajectory fills S almost without holes.

3.2 Anti-integrable limit Chaotic trajectories can be constructed analytically. We show how to do this in the anti-integrable limit, i.e., for large ε . First, we rewrite the dynamical equations (1) in the “Lagrangian form”. Let (xk , yk ), k ∈ Z be a trajectory of the standard map. Then for all integer k xk+1 = xk + yk + ε sin xk ,

yk+1 = yk + ε sin xk .

(2)

Eliminating the momenta yk , we get: xk+1 − 2xk + xk−1 = ε sin xk .

(3)

In fact, we have replaced the (non-compact) phase space Z by (compact) T2 , where T2 = {(x, y) mod 2π }.

6

10

D.V. Treschev

The map takes the form (xk−1 , xk ) → (xk , xk+1 ), and the phase cylinder becomes: {(x− , x) ∈ R2 }/ ∼, where the equivalence relation ∼ identifies any two points

, x ) and (x

, x

) such that (x− −

− x− = x − x

= 2π l, x−

l ∈ Z.

Now trajectories of the map are the sequences {xk }k∈Z , satisfying (3). In case of necessity yk can be calculated by using the first equation (2). Consider the case ε = ∞. Formally speaking, for ε = ∞ there is no dynamics: xk+1 can not be expressed in terms of xk−1 and xk . However still there are some “trajectories”. Indeed, dividing by ε , we obtain: 1 sin xk = (xk+1 − 2xk + xk−1 ) = 0. ε Hence, for ε = ∞ trajectories are sequences of the form xk = π lk ,

lk ∈ Z.

(4)

It turns out that for large ε the standard map has many trajectories similar to (4). Take a large positive number Λ and define the space of codes CΛ which consists of sequences a = {ak }k∈Z ,

ak = π lk ,

lk ∈ Z,

|ak+1 − ak |
Λ .

Hence CΛ is the space of sequences (4) such that the distances between the points ak+1 and ak are bounded from above by Λ . For any code a ∈ CΛ we define the metric space of sequences Πa : x = {xk }k∈Z ,

sup |xk − ak | < ∞. k∈Z

Metric on Πa has the form

ρ (x , x

) = sup |xk − xk

|,

x , x

∈ Πa .

k∈Z

Theorem 4 Given Λ > 0 and σ > 0 there exists ε0 = ε0 (Λ , σ ) > 0 such that for any code a ∈ CΛ and any ε > ε0 the standard map has a trajectory xˆ ∈ Πa with ρ (x, ˆ a) < σ . The trajectory x from Theorem 4 follows the code a in the sense that any point xk differs from ak not more than by σ . Hence, we have constructed a set of trajectories of the standard map which are in one-to-one correspondence with CΛ . Problem 10 What is the cardinality of CΛ ?

Some aspects of finite-dimensional Hamiltonian dynamics

11

It is natural to regard the trajectories xˆ as chaotic because according to our order they jump along σ -neighborhoods of the set π Z. In fact, there is a more serious motivation to say about chaos in this situation.7

3.3 Proof of the Aubry theorem The proof is based on the contraction principle in the metric space (Πa , ρ ). Equations (3) can be presented in the form xk = arcsink


x

k+1 − 2xk + xk−1

ε

 ,

(5)

where arcsink is the branch of arcsinus such that arcsink (0) = ak ∈ π Z. Hence arcsink maps the interval (−1, 1) onto the interval (ak − π2 , ak + π2 ), and the trajectory x = a satisfies (5) for ε = ∞. ˆ satisfying (5), as follows. Consider the map For big ε it is natural to construct x, x → x˜ = W (x) such that x˜k = arcsink


x

k+1 − 2xk + xk−1

ε

 .

Any fixed point of W is obviously a trajectory of the standard map. Lemma 3 Let ε > ε0 , where ε0 = ε0 (Λ , σ ) is sufficiently large. Then 1. W is defined on the ball Ba,σ ⊂ Πa with center a and radius σ ; 2. W (Ba,σ ) ⊂ Ba,σ ; 3. W is a contracting map on Ba,σ , i.e.,

ρ (W (x ),W (x

))


2(Λ + 2σ ) . sin σ

(3). Note that for any pair of real numbers u , u

∈ (− sin σ , sin σ ) | arcsink u − arcsink u

|
Here the multiplier

1 cos σ

=

1 |u − u

|. cos σ

d | du arcsink u|.

sup

u∈(− sin σ ,sin σ ) x˜

= W (x

). Then

We put x˜ = W (x ), for any k ∈ Z 
x − 2x + x 
x

− 2x

+ x

  k k−1 k k−1  |x˜k − x˜k

| =  arcsink k+1 − arcsink k+1  ε ε



 x

− 2xk

+ xk−1 − 2xk + xk−1 1  xk+1  − k+1
  cos σ ε ε





|x − xk+1 | + 2|xk − xk | + |xk+1 − xk+1 |
k+1 ε cos σ 4



ρ (x , x ).
ε cos σ Hence, inequality (6) holds if

ε0 >

8 . cos σ



3.4 Some remarks I would like to mention one unpleasant fact, which is that all methods that are known to date give a metrically negligible chaotic set in Tε and analogous systems. I mean the following. Given an arbitrary ε consider a set of chaotic trajectories that can be constructed by all methods, known by now. This subset of the cylinder Z has zero measure. This contradicts to our physical intuition, for large ε chaos should dominate. The results of computer simulations also show that this should be the case. But maybe we should not believe these computer pictures, as the precision of computations is necessarily finite. Nevertheless, most of specialists believe that the following conjecture is true. Conjecture. For ε = 0 in the standard map, chaos lives on sets of positive measure.

Some aspects of finite-dimensional Hamiltonian dynamics

13

4 Separatrix splitting 4.1 Poincaré’s observation Consider a Hamiltonian system, obtained as a non-autonomous perturbation of a system with one degree of freedom:

∂H , ∂y

y˙ = −

∂H , ∂x

(x, y) ∈ D ⊂ R2 .

(1)

H(x, y,t, ε ) = H0 (x, y) + ε H1 (x, y,t) + O(ε 2 ).

(2)

x˙ = Here D is a domain and

We assume that H is 2π -periodic in t and ε is a small parameter. Let z0 = (x0 , y0 ) ∈ D be an equilibrium in the unperturbed (ε = 0) system: grad H0 (z0 ) = 0. In the extended phase space D × T instead of the equilibrium we have the 2π -periodic solution z0 × T. Suppose that the equilibrium position (and therefore, the corresponding periodic solution) is hyperbolic. This means the following. Let ⎞ ⎛ 2 ∂ H0 ∂ 2 H0 ⎜ ∂ x∂ y ∂ y2 ⎟ ⎟ A=⎜ ⎝ ∂ 2 H0 ∂ 2 H0 ⎠ (z0 ) − − ∂ x2 ∂ y∂ x be the matrix determined by the linearization of (1)|ε =0 at z0 . Then tr A = 0. Hyperbolicity means that eigenvalues of A are outside the imaginary axis, i.e., det A < 0. Hyperbolic equilibria of Hamiltonian systems are exponentially unstable. On the critical energy level H0 (x, y) = H0 (z0 ) asymptotic curves (separatrices) Λ s,u are situated.8 We assume that the separatrices are doubled: Λ s = Λ u = Λ . In the extended phase space we have 2-dimensional asymptotic surfaces Λ s × T = Λ u × T = Λ × T. Problem 11 Prove that for small values of ε the perturbed system has a 2π -periodic solution (σε (t),t), σε (t) = z0 + O(ε ) ∈ D. The periodic solution (σε (t),t) is hyperbolic. Hence by the Hadamard-Perron theorem 9 there are surfaces Wεs,u ⊂ D × T, asymptotic to (σε (t),t). They are small deformations of the unperturbed surfaces W0s,u = Λ s,u × T. Poincaré discovered that Wεs and Wεu are generically distinct for ε = 0. Let us draw these surfaces. We will present a picture on the Poincaré section D × {0}. Hence, the periodic solution (σε (t),t) is presented by the point zε = σε (0), and instead of the surfaces Wεs,u we have the curves Λεs,u = Wεs,u ∩ {t = 0}. 8 9

s from “stable” and u from “unstable”: not very good, but traditional notation. Poincaré could prove this theorem, for analytic systems.

14

D.V. Treschev

Fig. 2 A complicated behavior of the separatrices for ε = 0 (right) unlike the unperturbed case (left) on the Poincaré section {(x, y,t) : t = 0 mod 2π }. The dashed domains are mapped by Tε to each other. Hence, their areas are the same

To obtain the right-hand part of Fig. 2, one should keep in mind the following: (a) For small ε the curves Λ u and Λεu (and also Λ s and Λεs ) differ just a little, at least, till Λεs,u are not far away from zε . (b) Λεs,u are invariant w.r.t. the Poincaré map Tε . (c) Λεs,u have no self-intersections, but can intersect each other. (d) Any intersection point z∗ = zε of the curves Λεs and Λεu (a homoclinic point) is mapped by Tε (and by Tε−1 ) to a homoclinic point. (e) Near the fixed point zε Tε is approximately determined by its linear approximation: it extends along Λεu and contracts along Λεs . (f) Tε and Tε−1 preserve area. Now it remains to assume that the curves Λεs and Λεu intersect transversally at some point z∗ , and the right-hand side of Fig. 2 readily appears. The complicated entangled net formed by the curves Λεs,u is an evidence of the complicated dynamics in the perturbed system.

4.2 The Poincaré integral To measure the separatrix splitting, we calculate the area of a lobe, presented in Fig. 2. The main tool for this and similar calculations is the Poincaré integral. Let γ (t) be the natural parametrization of Λ , i.e.,

γ (t) = (x(t), ˆ y(t)) ˆ

(3)

is a solution of (1). Since addition to the Hamiltonian of a function, depending only on t and ε , does not influence on the dynamics, we will assume that H1 (z0 ,t) ≡ 0.

Some aspects of finite-dimensional Hamiltonian dynamics

15

Then the Poincaré integral P(τ ) =

 +∞ −∞

H1 (γ (t + τ ),t) dt

converges. Problem 12 Prove that P(τ ) is 2π -periodic. Problem 13 Prove the identity dP(τ ) = dτ

 +∞ −∞

{H0 , H1 }(γ (t + τ ),t) dt,

The function P contains all information on the separatrix splitting in the first approximation in ε . Theorem 5 Let τ1 and τ2 be two neighboring non-degenerate critical points of P. Then there are two associated to them homoclinic points such that the area A (ε ) of the corresponding lobe equals A (ε ) = |ε P(τ1 ) − ε P(τ2 )| + O(ε 2 ).

(4)

4.3 Proof of Theorem 5 4.3.1 Hamilton–Jacobi equation Following Poincaré, consider the case when Λ projects one-to-one to the axis. In the general case the proof is based on the same ideas. The curve Λ (see Fig. 3) can be determined by the equation y = ∂∂ϕx (x) for some function ϕ (x). We have an analogous equation in the extended phase space, i.e., the surface W0s = W0u has the form ∂ϕ (x) . (x, y,t) : y = ∂x

Fig. 3 The case, considered in the proof of Theorem 5, appears when x is an angular variable. For example, for non-autonomous perturbation of a pendulum. The corresponding separatrices look as in the figure

16

D.V. Treschev

The perturbed asymptotic surfaces are as follows:

∂ Ss,u (x,t, ε ) , (x, y,t) : y = ∂x

Ss,u (x,t, 0) = ϕ (x).

Remark 2 The functions Ss,u are defined non-uniquely: up to an addition of arbitrary functions f s,u (t, ε ). Proposition 4 One can assume that Ss,u satisfy the Hamilton-Jacobi equation 
∂ Ss,u ∂ Ss,u (x,t, ε ) + H x, (x,t, ε ),t, ε = 0. ∂t ∂x

(5)

Remark 3 Equation (5) for ε = 0 shows that if we want the equations Ss,u (x,t, 0)= ϕ (x) to hold exactly (not up to an addition of a function of t), we should put H0 |Λ =0. Proof of Proposition 4 is based on a direct calculation. Let (x, y,t) = (x, ∂∂ Sx (x,t, ε ),t) be a point, lying on Wε (for brevity we do not write the indices s, u), and ()· = dtd , denotes the time derivative w.r.t. equations (1). Then

∂ 2S ∂ 2S (x,t, ε ) + 2 (x,t, ε ) x˙ ∂ x∂ t ∂x ∂H (x, y,t, ε ) =− ∂x  ∂H ∂
∂S ∂ 2S (x, y,t, ε ) 2 (x,t, ε ). = − H x, (x,t, ε ),t, ε + ∂x ∂x ∂y ∂x

y˙ =

Since

∂ 2S ∂ x2

x˙ =

∂ H ∂ 2S ∂ y ∂ x2 ,

we get:


∂S  ∂ ∂S (x,t, ε ) + H x, (x,t, ε ),t = 0. ∂x ∂t ∂x

Hence for some function α (t, ε )
∂S  ∂S (x,t, ε ) + H x, (x,t, ε ),t = α (t, ε ). ∂t ∂x By Remark 2 α can be taken equal to zero.



s,u

4.3.2 The function S1s , u and the Poincaré integral Expand equations (5) in power series in ε . Let S = ϕ (x) + ε S1 (x,t) + O(ε 2 ). In zero approximation we have:
∂ϕ  ∂ϕ (x) + H0 x, = 0. ∂t ∂x (compare with Remark 3).

Some aspects of finite-dimensional Hamiltonian dynamics

17

The first approximation is as follows:
∂ ϕ  ∂ H
∂ ϕ  ∂ 2 Ss,u ∂ S1s,u 0 1 (x,t) + H1 x, ,t + x, (x,t) = 0. ∂t ∂x ∂y ∂ x ∂ x∂ t

(6)

Since ∂ H0 /∂ y = x, ˙ equation (6) can be rewritten in the form
∂ϕ  d s,u S1 (x,t) + H1 x, ,t = 0. dt ∂x

(7)

Plugging in (7) instead of x its parametrization x(t ˆ + τ ) (see (3)), we get: d s,u S (x(t ˆ + τ ),t) = H1 (γ (t + τ ),t). dt 1 Now we integrate in t: S1s (x(t ˆ + τ ),t) − S1s (x(+∞),t) ˆ =

 +∞ t

S1u (x(t ˆ + τ ),t) − S1u (x(−∞),t) ˆ =−

H1 (γ (s + τ ), s) ds,

 t −∞

H1 (γ (s + τ ), s) ds.

(Recall that x(−∞) ˆ = x(+∞) ˆ = x0 .) Hence ˆ + τ ),t) − S1u (x(t ˆ + τ ),t) = P(τ ) + β (t), S1s (x(t where β (t) = S1s (x0 ,t) − S1u (x0 ,t). Differentiating in τ , we get:
 ˙ˆ + τ ) ∂ S1s (x(t ˆ + τ ),t) − S1u (x(t ˆ + τ ),t) = P (τ ). x(t ∂x

(8)

4.3.3 Homoclinic points and lobes Homoclinic points are determined by the equations (x, ∂∂Sx ) = (x, ∂∂Sx ) i.e., s

u

∂ Ss ∂ Su ∂ϕ ∂ϕ (x(t ˆ + τ )) + ε 1 (x(t ˆ + τ ),t) − (x(t ˆ + τ )) − ε 1 (x(t ˆ + τ ),t) + O(ε 2 ) = 0, ∂x ∂x ∂x ∂x where we take again x(t ˆ + τ ) instead of x. According to (8) and the relation ˙ˆ + τ ) = 0 we get: x(t P (τ ) + O(ε ) = 0. Hence non-degenerate critical points of P(τ ) generate homoclinic points. Question. Why we need non-degeneracy and in what sense we use the word “generate”?

18

D.V. Treschev

Let τ1 and τ2 be two consecutive non-degenerate critical points of P(τ ). The corresponding homoclinic points z1 = (x1 , y1 ), z2 = (x2 , y2 ) on the Poincaré section {t = 0 mod 2π } are “angles” of a lobe. Let A (ε ) be its area. Then  x
s    2 ∂S ∂ Su (x, 0, ε ) − (x, 0, ε ) dx A (ε ) =  ∂x ∂x x   1x
   2 ∂ S1s Su ∂ (x, 0) − ε 1 (x, 0) dx + O(ε 2 ). ε =  ∂x ∂x x 1

We change variables x = x( ˆ τ ) in the integral and use (8):  τ    2 ∂
s  u  ˙ˆ τ ) d τ  + O(ε 2 ) S1 (x( ε ˆ τ ), 0) − S1 (x( ˆ τ ), 0) x( A (ε ) =   ∂x τ1   τ2    = ε P (τ ) d τ  + O(ε 2 ). τ1

This implies (4).



4.4 Standard example Consider a pendulum with a vertically oscillating suspension point, i.e., the system with Hamiltonian 1 H(x, y,t, ε ) = y2 + Ω 2 cos x + εθ (t) cos x. 2

(9)

Performing in case of necessity the change t → λ t, we can assume that θ is 2π periodic. A natural parametrization on the unperturbed separatrix γ (t) can be computed explicitly. Problem 14 Check that cos(x(t)) ˆ = 1 − 2 cosh−2 (2Ω t). Hence P(τ ) =

 +∞ −∞

θ (t)(cos(x(t ˆ + τ )) − 1) dt.

Problem 15 Check that for θ (t) = cost P(τ ) = −

π cos τ

. 2Ω 2 sinh( 2πΩ )

If θ (t) = cost lobes have the areas A (ε ) =

επ + O(ε 2 ). Ω 2 sinh( 2πΩ )

Some aspects of finite-dimensional Hamiltonian dynamics

19

References 1. Abraham, R, Marsden, J. Foundations of mechanics. Benjamin/Cummings Publishing Co., Inc., Advanced Book Program, Reading, MA, 1978. 2. Arnold, V. I. Mathematical methods of classical mechanics. Graduate Texts in Mathematics, 60. Springer, New York, 1989. 3. Arnold, V. I., Kozlov, V. V., Neishtadt, A. I. Mathematical aspects of classical and celestial mechanics. Encyclopaedia Math. Sci., 3, Springer, Berlin, 1993. Springer, Berlin, 1997.

Four lectures on the N-body problem Alain Chenciner1

Abstract In the first two lectures, Hamiltonian techniques are applied to avatars of the N-body problem of interest to astronomers: the first one introduces one of the simplest non integrable equations, the planar circular restricted problem in the lunar case, where most degeneracies of the general (non-restricted) problem are not present; the second one is a quick introduction to Arnold’s theorem on the stability of the planetary problem where degeneracies are dealt with thanks to Herman’s normal form theorem. The last two lectures address the general (non-perturbative) N-body problem: in the third one, a sketch of proof is given of Marchal’s theorem on the absence of collisions in paths of N-body configurations with given endpoints which are local action minimizers; in the last one, this theorem is used to prove the existence of various families of periodic and quasi-periodic solutions with prescribed symmetries and in particular to extend globally Lyapunov families bifurcating from polygonal relative equilibria. Celestial mechanics is famous for demanding extensive computations which hardly appear here: these notes only describe the skeleton on which these computations live.

1 The Poincaré–Birkhoff–Conley twist map of the annulus for the planar circular restricted three-body problem 1.1 The Kepler problem as an oscillator The (normalized) motions in a plane of a particle submitted to the Newtonian attraction of a fixed center – the so-called Kepler problem – are the solutions of the equation x¨ = −x/|x|3 , 1

University Paris 7 and IMCCE (Paris Observatory) e-mail: [email protected]

W. Craig (ed.), Hamiltonian Dynamical Systems and Applications, 21–52. c 2008 Springer Science + Business Media B.V.


21

22

A. Chenciner

where x ∈ R2 = C is identified with a complex number and the dot denotes the time derivative. These equations are the Hamilton equations x˙ =

∂H ∂H , y˙ = − ∂ y¯ ∂ x¯

associated to the Hamiltonian H : (C \ {0}) × C → R and the symplectic form ω respectively defined by H(x, y) = |y|2 − 2/|x|,

ω = dx ∧ d y¯ + d x¯ ∧ dy.

The Levi-Civita mapping (z, w) → (x = 2z2 , y = w/ε z¯) defines a two-fold covering (L.C.)

K −1 (0) \ {z = 0} → Σε = H −1 (−1/ε 2 )

from the complement of the plane z = 0 in the 0-energy three-sphere K −1 (0) of the harmonic oscillator     K(z, w) = |z|2 + |w|2 − ε 2 = ε 2 |z|2 H 2z2 , w/ε z¯ + 1/ε 2 , to the energy hypersurface Σε = H −1 (−1/ε 2 ) of the Kepler problem (both diffeomorphic to S1 × R2 ). It is conformally symplectic and sends integral curves of the harmonic oscillator with energy ε 2 to those of the Kepler problem with energy −1/ε 2 after the change of time dt = 2ε |x|dt which prevents the velocity to become infinite at collision. In the coordinates u1 = w + iz, u2 = w¯ + i¯z these integral curves are u1 (t) = c1 eit , u2 (t) = c2 eit , |c1 |2 + |c2 |2 = 2ε 2 , that is the intersections of the three-sphere with the complex lines u1 /u2 = cste, or in other words the fibers of the Hopf fibration (u1 , u2 ) → u1 /u2 : S3 → P1 (C). The closest approximation to a section of the Hopf map, the annulus arg u1 + arg u2 = 0 (mod 2π ) is a global surface of section of the flow of the Harmonic oscillator in a sphere of constant energy: with the exception of the two fibers which form its boundary, all the fibers cut this annulus transversally in two points; hence, the second return map is the identity. Thus perturbations of the Kepler problem with negative energy are essentially perturbations of the identity map. This is one of the main sources of degeneracies in celestial mechanics.

1.2 The restricted problem in the lunar case The equations of the N-body problem m j (r j − ri ) 3 j =i ||ri − r j ||

r¨i = g ∑

Four lectures on the N-body problem

23

make sense even if some of the masses vanish. Such masses are influenced by the non-zero masses but do not influence them. We shall consider two primaries, say the Sun (mass µ ) and the Earth (mass ν ) which have a uniform circular motion around their center of mass and a zero-mass third body, say the Moon, which stays close to the Earth. We shall use the normalization g = 1 and µ + ν = 1. We identify the inertial plane with C (coordinate X = X1 + iX2 centered on the center of mass of the couple Sun-Earth) and introduce a rotating complex coordinate x = x1 + ix2 = Xe−iω t − µ centered on the Earth. Setting y = x˙ + iω x (up to a translation, this is the velocity in the inertial frame), the equations of motion of the Moon take the Hamiltonian form ∂H ∂H , y˙ = − , x˙ = ∂ y¯ ∂ x¯ where H is the Jacobi integral (the constant 2µ is added for convenience) H(x, y) = |y|2 + iω (xy ¯ − xy) ¯ −

2µ 2ν − − µ (x + x) ¯ + 2µ . |x| |x + 1|

More precisely, the vector field is the symplectic gradient of the symplectic form

ω = dx ∧ d y¯ + d x¯ ∧ dy = 2(dx1 ∧ dy1 + dx2 ∧ dy2 ). As in the first section, we consider the energy hypersurface H −1 (1/ε 2 ), with ε a small parameter. Its projection on the x plane is made of three connected components: a neighborhood of the Sun, a neighborhood of the Earth and a neighborhood of infinity (the so-called Hill’s regions, which imply Hill’s stability result, praised by Poincaré). We shall be interested in the connected component of H −1 (1/ε 2 ) where |x| stays small. Then   1 2 3 2 2ν 2

2 − 2y µ |x| + (x + x¯ ) + O3 (x) . ¯ − xy) ¯ − H(x, y) = |y| + iω (xy |x| 4 8 We see that the influence of the Sun on the Moon becomes negligible with respect to the one of the Earth and that at the collision limit, it disappears and one is left with a Kepler problem. To make this apparent, we again apply the Levi–Civita transformation. We get     1 1 2 2 2 w + H 2z = f 2 (z, w)|z|2 + |w|2 − νε 2 − ε 2 µ g(z), ε |z| , K(z, w) = ε z¯ ε2 2 where  f (z, w) = 1 + 2iε (¯zw − zw), ¯

 2

g(z) = 2|z|

 1 2 2 − 1 + z + z¯ . |2z2 + 1|

As in the Kepler case, the direct image of the restriction to K −1 (0) \ {z = 0} of the Hamiltonian flow z˙ = ∂∂ Kw¯ , w˙ = − ∂∂Kz¯ becomes the flow of the restricted problem with Jacobi constant −1/ε 2 after the change of time dt = 2ε |x|dt .

24

A. Chenciner

Each truncation of the Taylor expansion of K(z, w) at the origin, ¯ ε 2 µ (2|z|6 +3|z|2 (z4 + z¯4 )+08 (z)), K(z, w) = −νε 2 +|z|2 +|w|2 +2iε |z|2 (¯zw− wz)− makes sense dynamically when restricted to K −1 (0) : we get At order 2, the harmonic oscillator, which regularizes the Kepler problem At order 4, the regularization of the Kepler problem in a rotating frame At order 6, Hill’s problem. This is the highest order of interest to us

1.3 Hill’s solutions ˆ w) = −νε 2 + f 2 (z, w)|z|2 + w2 of K at fourth order is a comThe truncation K(z, pletely integrable Hamiltonian, a first integral being the angular momentum or, what is equivalent, the function f 2 (z, w). This is not surprising as we already knew that the restriction to K −1 (0) corresponds to the completely integrable Kepler problem in a rotating frame. The intersection of level hypersurfaces of K and f 2 defines in general a two-dimensional torus, except when the two hypersurfaces are tangent, that is when w = ±i f (z, w)z. In this case the intersection degenerates to a circle; in K −1 (0), this defines two solutions which project (by a 2-1 map) onto the two circular solutions (one direct, one retrograde) of the rotating Kepler problem with the given value −1/ε 2 of the Jacobi constant. From now on, two roads may be followed: one can, along with Kummer [Ku], stick to symplectic coordinates or one can, as did Conley, use the simpler but not symplectic coordinates

ξ1 = w + i f (z, w)z, We shall follow Conley. The equations z˙ =

ξ2 = w¯ + i f (z, w)¯z. ∂K ˙ = − ∂∂Kz¯ ∂ w¯ , w

take the form


 ε ξ˙1 = iξ1 1 − |ξ1 − ξ¯2 |2 + ε 2 O5 (ξ1 , ξ2 ), 2
 ε ˙ ξ2 = iξ2 1 + |ξ1 − ξ¯2 |2 + ε 2 O5 (ξ1 , ξ2 ). 2 For this section, we do not need the exact expression of the terms of order 5. We shall show that the energy hypersurface K −1 (0) contains two periodic solutions of minimal periods close to 2π , corresponding to the so-called Hill’s lunar orbits, direct and retrograde, which are almost circular periodic motions of the Moon around the Earth in the rotating frame. The value 0 of the energy does not play a special role and it is in fact possible to prove the existence of two “Lyapunov” families of periodic solutions stemming from the origin and foliating two smooth (even analytical) germs of invariant surfaces in the (z, w) four-dimensional phase space. This is a degenerate version of Lyapunov’ theorem, the degeneracy being the double eigenvalues ±i of the linearization ξ˙1 = iξ1 , ξ˙2 = iξ2 , of the vector-field at

Four lectures on the N-body problem

25

ξ1 = ξ2 = 0. Recall that this degeneracy comes from the fact that all solutions of the Kepler problem with a given energy are periodic with the same period. Here are the main steps of the proof of the existence of Hill’s orbits. 1. Putting the vector-field into normal form at order 3: the idea, which goes back to Poincaré’s thesis and was much developed by Birkhoff, is to simplify as much as possible a finite part of the vector-field’s Taylor expansion at the origin by means of local change of variables tangent to Identity. It relies on the fact that replacing X = (x1 , · · · , xn ) by Y = X + h(X), where the components of h(X) start with terms homogeneous in X of degree r, transforms the equation X˙ = AX + F(X) into the equation Y˙ = AY + [A, h](Y ) + Or+1 , where [, ] is the Lie bracket of the two vector-fields. If A = diag(λ1 , · · · , λn ) and h = (h1 , . . . , hn ) with hs (Y ) = yi11 · · · yinn and h j = 0 if j = s, one checks that [A, h] = k with ks (Y ) = (i1 λ1 + · · · + in λn − λs )yi11 · · · yinn and k j = 0 if j = s. It follows that one can suppress only non-resonant terms, i.e. those for which no resonance relation i1 λ1 + · · · + in λn − λs = is satisfied. In our case, this allows to replace the equations by the following (we kept the same name for the variables):   ξ˙1 = iξ1 1 + α |ξ1 |2 + β |ξ2 |2 + ε 2 ϕ1 (ξ1 , ξ2 ),   ξ˙2 = iξ2 1 + a|ξ1 |2 + b|ξ2 |2 ) + ε 2 ϕ2 (ξ1 , ξ2 ), with α = β = − ε2 , a = b = + ε2 , ϕ1 and ϕ2 of order 5 in ξ1 , ξ2 , ξ¯1 , ξ¯2 . In the neighborhood of the origin, the flow Φt (ξ1 , ξ2 ) = (ξ1 (t), ξ2 (t)) can be written   ξ1 (t) = eit ξ1 (1 + i(α |ξ1 |2 + β |ξ2 |2 )t) + ε 2 α1 (ξ1 , ξ2 ,t) ,   ξ2 (t) = eit ξ2 (1 + i(a|ξ1 |2 + b|ξ2 |2 )t) + ε 2 α2 (ξ1 , ξ2 ,t) , with α1 , α2 of order 5 in ξ1 , ξ2 , ξ¯1 , ξ¯2 uniformly in t belonging to a compact. 2. Regularizing the equations for a periodic solution by means of a blow-up: We look for a periodic solution whose period T is close to the period 2π of the solution ξ2 = 0 of the rotating Kepler problem approximation (an analogous reasoning can be made for a solution close to ξ1 = 0). Because of the existence of the energy first integral, the equations which define a periodic solution of period T , that is ξ1 (T ) = ξ1 , ξ2 (T ) = ξ2 , are consequence of the equations Arg ξ1 (T ) − Arg ξ1 = 2π ,

ξ2 (T ) − ξ2 = 0.

Writing down directly these equations would lead to possibly non differentiable terms like α1 (ξ1 , ξ2 )/ξ1 . Indeed, they read   α1 (ξ1 , ξ2 , T ) , 2π = T + arg 1 + i(α |ξ1 |2 + β |ξ2 |2 )T + ε 2 ξ1  iT    e 1 + i(a|ξ1 |2 + b|ξ2 |2 )T − 1 ξ2 + ε 2 eiT α2 (ξ1 , ξ2 , T ) = 0.

26

A. Chenciner

We solve this problem by a further localization in a domain of the form |ξ2 |
|ξ1 | by means of a complex blow-up

ξ1 = z 1 ,

ξ2 = z 1 z 2

which replaces such a term by α1 (z1 , z1 z2 )/z1 which is now differentiable. The first equation determines T as a C3 function of z1 , z¯1 , z2 , z¯2 , T = 2π − 2π |z1 |2 (α + β |z2 |2 ) + o3 , where o3 vanishes at order 3 along z1 = 0. The second one becomes 2π i|z1 |2 (a − α + (b − β )|z2 |2 )z2 + o3 = 03 . As a − α = ε = 0, solving this equation leads to a C1 surface tangent to the plane z2 = 0, that is in the (ξ1 , ξ2 ) space to a C2 surface N1 tangent at order 2 to the plane ξ2 = 0. Intersecting with the energy hypersurface K = 0 gives the seeked for periodic solution. In the same way, one proves the existence of N2 tangent to ξ1 = 0. 3. Proving the analyticity of N1 and N2 : This is done in Conley’s thesis by closely following the proof given in the non-resonant case by Siegel and Moser. To understand the formulas, one suppresses the resonant terms of any order by means of a formal (not convergent !) transformation. One gets new (formal coordinates) ζ1 , ζ2 such that ζ˙1 and ζ˙2 become formal series in the resonant terms ζi |ζ j |2 and ζi (ζ j ζ¯k ). Rewriting the computation of periodic solutions as above leads to formal surfaces N1 and N2 where, for example, N1 is defined by a (formal) equation of the form ζ2 = γ (|ζ1 |2 )ζ1 , the restriction of the vectorfield being of the form ζ˙1 = α (|ζ1 |2 )ζ1 where α has purely imaginary values (this corresponds to the fact that N1 is foliated by periodic solutions surrounding the origin). One proves the convergence of γ and α by writing down majorant series.

1.4 The annulus twist map Replacing the boundaries ξ1 = 0 and ξ2 = 0 of the Kepler annulus by the two Hill orbits, one can now construct a global annulus of section of the flow in the threesphere K −1 (0) and analyze the first return map. Such an annulus is of course not unique and it will be convenient to chose it so as to contain the “collision circle” of equation z = 0. In order to get precise enough information on the first return map, one must analyze the equations up to the 5th order where the influence of the Sun comes into play. Writing down a normal form up to this order implies first computing the effect on terms of order five of the change of variables leading to a normal form at order 3. In fact, one can dispense with this: it is enough to suppress only the

Four lectures on the N-body problem

27

non resonant terms of order 5, keeping the terms of order 3 as they stood initially. Moreover, the above analysis of the submanifolds N1 and N2 whose intersection with K = 0 defines Hill’s orbits, shows that there exists an analytic change of variables which transforms them into coordinate planes. A finer analysis shows that such a straightening change of variables differs from Id only by terms ε A + ε 2 B, where A is resonant of order 5 and B is of order 7. One deduces that such a straightening of N1 and N2 does not bring any new change to the differential equation up to order 5. Finally, we get new coordinates (ζ1 , ζ2 ) such that N1 and N2 are respectively defined by ζ1 = 0 and ζ2 = 0, and the energy hypersurface K −1 (0) and the collision circle z = 0 by 1 (|ζ1 |2 + |ζ2 |2 ) − νε 2 + ε O6 (ζ ) = 0, 2

and

ζ1 − ζ¯2 + ε O5 (ζ ) = 0.

It follows that an annulus of section in K −1 (0) containing the collision circle and bounded by the Hill orbits can be defined by the equation Arg ζ1 + Arg ζ2 + ε O4 (ζ ) = 0

(mod 2π ).

Computing a little more, one can find coordinates (ϕ , ρ ) on this annulus, such that the two boundaries are close to ρ = ±1 and the first return map takes the form   µ 1 ν 3ν 2 (1 − )ε 6 ρ + 0(ε 7 ), ρ + O(ε 7 ) . Pε (ϕ , ρ ) = ϕ + − ε 3 − 2 2 2 4 Coming back to the definition of this annulus, one checks that the return map corresponds essentially to the passages of the orbit of the Moon through aphelium in the rotating frame. Originating from a Hamiltonian system, this map necessarily preserves a measure defined by a smooth density. Moreover, it is a O(ε 7 ) perturbation of an integrable twist map whose twist is of size ε 6 . This is a perfect ground for applying the main results of the general theory of conservative twist maps, a particular case of the theory of Hamiltonian systems with two degrees of freedom: 1. Applied to the iterates of the return map, the Birkhoff fixed point theorem yields an infinite number of periodic orbits of higher and higher periods to which correspond periodic orbits of long period of the Moon around the Earth in the rotating frame. 2. The Moser invariant curve theorem implies the existence of a positive measure Cantor set of invariant curves on which the map is conjugated to a diophantine irrational rotation and to which correspond quasi periodic orbits of the Moon. 3. To the Liouville rotation numbers, the Aubry–Mather theory associates invariant Cantor sets to which correspond orbits of the Moon with a Cantor caustic. 4. Finally, it is possible to prove that the image of the collision circle intersects itself transversally at eight points [CL]; in particular, it is not contained in an invariant curve. Varying the value of ε moves the invariant curve of a given rotation number across the annulus which forces intersection with the collision curve. This proves the existence of invariant “punctured” tori which correspond

28

A. Chenciner

to orbits of the Moon which persistently change their direction of rotation around the Earth in the rotating frame (generalization of the punctured tori to the full planar three-body problem were given by Féjoz in his thesis [Fe1]). Remark. For writing down formulas, working in the two-fold covering K −1 (0) of the energy hypersurface diffeomorphic to S3 is convenient but one can prefer to state the results downstairs in the compactification (regularization), diffeomorphic to SO(3) (that is to the real projective space of dimension 3), of the original energy hypersurface H −1 (− ε12 ). The first return map then becomes a perturbation of the Identity (the Kepler case) of the form
 µ Pε (ϕ˜ , ρ ) = ϕ˜ − νε 3 − 3ν 2 (1 − )ε 6 ρ + 0(ε 7 ), ρ + O(ε 7 ) . 4 and the collision curve intersects its image only four times. A problem. When the collision curve intersects the set of invariant curves, the closure of the union of its iterates, containing the set of intersected curves, is of positive measure. What if the collision curve is contained in a Birkhoff region of instability? X x

E M

O S

Fig. 1 Hill’s regions

x

Hill’s region

X

Four lectures on the N-body problem

29

Fig. 2 The annulus of section

2 The Arnold–Herman stability theorem for the spatial (1 + n)-body problem In the so-called planetary problem, one mass m0 is dominant (the Sun) and the others, the planets are of the form ε m1 , . . . , ε mn , where ε is small (around 10−3 for the “real” solar system). If x0 = (x01 , x02 , x03 ), x1 , . . . , xn ∈ R3 are the positions and ||.|| the euclidean norm, Newton’s equations read x¨ j = m0

x0 − x j xk − x j + ε ∑ mk , j = 1, . . . , n. 3 ||x0 − x j ||3 ||x k − x j || k = j

The solutions are the projections on the configuration space of the integral curves of the Hamiltonian vector field defined in the phase space, whose coordinates are denoted by (x0 , . . . , xn , y0 , ε y1 , . . . , ε yn ) and symplectic form is ∑1
k
3 dx0k ∧ dyk0 + ε ∑1
j
n ∑1
k
3 dxkj ∧ dykj , by the Hamiltonian   ||y j ||2 mo m j m j mk 1 ||y0 ||2 1 − ε2 ∑ · +ε − ∑ 2 m0 2 1
∑ m ||x − x || ||x j j j − xk || 0 j
n 1
j
n 1
j 0 or Λ˙ (t) < 0, then |Λ˙ (t)| defines a Euclidean structure on Λ (t) and RΛ (t) is a self-adjoint operator for this Euclidean structure. In particular, the operator RΛ (t) is diagonalizable and all its eigenvalues are real. We say that the curvature is positive or negative e. t. s. if all eigenvalues are like that. Set ν = sign(Λ˙ (t)). The matrix of the operator RΛ (t) is equal to the matrix of the quadratic form ν Λ˙ (t) in the coordinates where the Euclidean structure |Λ˙ (t)| is presented by the unit matrix. In particular, assumptions of Theorem 2 are satisfied if the quadratic form Λ˙ (t) is sign-definite and RΛ (t)
0. It is important that the construction of the curvature is intrinsic and thus survives under symplectic transformations. In other words, if A is a linear symplectic transformation of Σ and AΛ : t → A(Λ (t)), then the operator RAΛ (t) is similar to the operator RΛ (t) and has the same eigenvalues. Definition 5.1. Let ΛzΛ (·), z ∈ T ∗ M, be the Jacobi curves of the Hamiltonian field de f

H; then Rλz (0) = RH z is called the curvature operator of H at z. Recall that Λ˙ z (0) = − ∂∂ pH2 (z), z = (p, q). If H is strongly convex with respect to p, then Λ˙ z (0) < 0, ∀z. It was proved in Lecture 3 that the inequality Λ˙ z (0)
0, ∀z implies Λ˙ z (t)
0, ∀t. We can repeat that proof and see that the result remains valid if one substitute the non-strong inequality by the strong one. In fact, we can see much more if we analyze the proof. Indeed, the germ at t of the curve Λz (·) is the image of the germ at 0 of the curve ΛetH (z) (·) under the fixed symplectic transformation e−tH ∗ . Hence all invariant quantities of these germs are equal. In particular, the operator RΛz (t) has the same eigenvalues as the curvature operator of H at the point etH (z). This fact is very advantageous because the curvature of H is just a (rather complicated but quite explicit) differential operator of H, in particular we do not need to solve differential equations in order to compute this curvature. 2

Hamiltonian systems and optimal control

155

Combining all together, we obtain Theorem 5.2. If the restrictions of H to the fibers Tq∗ , q ∈ M, are strongly convex tH do not have functions and RH z
0, ∀z ∈ M, then the trajectories of the flow e conjugate points and, moreover, this flow possesses two invariant Lagrangian distributions Λz (±∞), z ∈ T ∗ M. Let us discuss what happens when RΛ  0. First of all, let us see how changes the curvature if we re-parameterize the curve Λ (t). Remark. Clearly, the presence of conjugate point does not depend on the parameterization. Let ϕ : Rn → Rn be a change of parameter; we assume that ϕ˙ (t) = 0. Denote 2 ... 1 3 Rϕ (t) = ϕ˙ −1 (t) ϕ (t)(t) − ϕ˙ −1 ϕ¨ (t) , 2 4 the Schwartzian derivative of ϕ (t). Let

Λϕ : t → Λ (ϕ (t)) be the re-parameterized curve. Proposition 5 RΛϕ (t) = ϕ˙ (t)2 RΛ (ϕ (t)) + Rϕ (t)Id. This formula (the chain rule) can be checked by direct calculation, which we omit here. √ Example. Take ϕ (t) = √1c arctan( ct). Then: Rϕ (t) =

−c , (ct 2 + 1)2

RΛϕ (t) =

(ct 2 + 1)2

1

ϕ˙ (t) =

1 , ct 2 + 1

(RΛ (ϕ (t)) − cId) .

Theorem 5.3 (Comparison theorem). Let t → Λ (t) be a smooth strongly monotone curve in the Lagrange Grassmannian and c be a nonnegative constant. 1. If RΛ (t)
cId, ∀t, then any pair of conjugate points t1 and t2 satisfies the inequality |t1 − t2 |  √πc . " ! 2. If 1n traceRΛ (t)  c, then any segment t,t + √πc contains a point conjugate to zero. Remark. If c → 0 in statement 1, then |t1 − t2 | → ∞ which correlates with our previous result. Proof. Statement 1. Re-parameterization and the chain rule reduces everything to the case of non-positive curvature (see above Example).

156

A. Agrachev

Statement 2. Assume that Λ (t) is transversal to some fixed Lagrangian subspace (for instance, to Λ (0)) for any t ∈ [t1 ,t2 ]. Then the segment Λ (t), t ∈ [t1 ,t2 ] of the curve is contained in the fixed coordinate chart of the Lagrange Grassmannian and can be presented in the matrix form:

Λ (t) = {(p, St p) : p ∈ Rn },

t ∈ [t1 ,t2 ].

Hence

... 1 3 RΛ (t) = S˙t−1 S t − (S˙t−1 S¨t )2 . 2 4 Now set Wt = 12 S˙t−1 S¨t ; we have, RΛ (t) = W˙ (t) − W (t)2 . What remain is to find the lower bound for the blow-up time of the solutions to the differential inequality traceW˙  trace(W 2 ) + nc. This is an easy task due to the fact that (traceW )2
n trace(W 2 ). In order to conclude we recall that in the Riemannian case the Hamiltonian flow is the geodesic flow. Actually, the Riemannian structure identifies T M and T ∗ M. 2 ∗ So, in this case ∂∂ pH2 is the Riemannian metric. The curvature RH z , where z ∈ Tq , is essentially the sectional curvature at q in the two-dimensional directions which include z ∈ Tq∗ M ∼ = Tq M.

References 1. Agrachev, A.A.: Geometry of optimal control problems and Hamiltonian systems. In: C.I.M.E. Lecture Notes in Mathematics, Springer, Berlin (to appear) 2. Agrachev, A.A., Sachkov, Yu.L.: Control Theory from the Geometric Viewpoint. Springer, Berlin (2004) 3. Gamkrelidze, R.V.: Principles of Optimal Control Theory. Plenum Publishing, New York (1978) 4. Jurdjevic, V.: Geometric Control Theory. Cambridge University Press, Cambridge (1997) 5. Pontryagin, L.S., Boltyanskij, V.G., Gamkrelidze, R.V., Mishchenko, E.F.: The Mathematical Theory of Optimal Processes. Pergamon, Oxfrod (1964)

KAM theory with applications to Hamiltonian partial differential equations Xiaoping Yuan1

Abstract In these notes I present a KAM theorem on the existence of lower dimensional invariant tori for a class of nearly integrable Hamiltonian systems of infinite dimensions, where the second Melnikov’s conditions are completely eliminated and the algebraic structure of the normal frequencies is not required. This theorem can be used to construct invariant tori and quasi-periodic solutions for nonlinear wave equations, Schrödinger equations and other equations of any spatial dimensions.

1 Brief history and basic ideas of KAM theory These lecture notes present a KAM theorem with applications to some nonlinear partial differential equations, such as nonlinear wave equations and Schrödinger equations of higher spatial dimensions. Although it is a powerful tool in dynamical systems, the KAM technique is usually thought to be very complicated, even tedious. I will omit some unimportant details so that the basic idea of the KAM theory can be clearly understood. Before stating the KAM theorem, let us recall some of the background, taking the nonlinear wave (NLW) equation as an example. One wants to find a periodic solution of NLW equation utt − uxx +V (x)u + g(x, u) = 0,

(1)

subject to Dirichlet boundary condition u(t, 0) = u(t, π ) = 0, where g is a nonlinear term. In the 1970s using variational methods, Rabinowitz [19] showed that there is a non-constant T -periodic solution u(t, x) ∈ L p (R × [0, π ]) if T /π ∈ Q, where p > 2 1

School of Mathematical Sciences, Fudan University, Shanghai 200433, China, Key Laboratory of Mathematics for Nonlinear Science (Fudan University), Ministry of Education e-mail: [email protected]; [email protected]

W. Craig (ed.), Hamiltonian Dynamical Systems and Applications, 157–177. c 2008 Springer Science + Business Media B.V.


157

158

X. Yuan

depends on g. The condition T /π ∈ Q guarantees some kind of compactness which is usually required in variational methods. A natural question is  What happens when T /π ∈ R \ Q? From a geometric viewpoint, a periodic solution can be regarded as an invariant torus of 1 dimension, i.e., an invariant closed curve, in some phase space. Thus, another natural question should be  Are there invariant tori of N-dimension with N > 1? If there are such tori, and any motion on the tori is quasi-periodic, it follows that there are time quasi-periodic solutions of (1). The previous questions can not be answered at the present time entirely by variational methods, since the compactness conditions can not be fulfilled. Fortunately, KAM theory can answer these questions. In order to see how the KAM theory adresses them, we write the equation (1) in a discrete form. To this end, we let λ j2 and φ j (x) be the eigenvalues and eigenfunctions, respectively, of the Sturm– Liouville problem1 −

d2y +V (x)y = λ y, y(0) = y(π ) . dx2

Note that {φ j (x) : j = 1, 2, ...} is a complete orthogonal system of L2 ([0, π ]). For simplicity we assume the nonlinearity g(x, u) = u3 without loss of generality. Since we will search for solutions of small amplitude, we can assume g = ε u3 where ε is a small parameter. This can be fulfilled by substituting Let v = ut . Then (1) reads

√ ε u for u in (1).

ut = v, vt = utt = −[−uxx +V (x)u + g(x, u)] Substituting for u and v the expressions ∞

u=

qj ∑ √λ φ j (x), v = j j=1

∞

∑



λ j p j φ j (x)

j=1

we get a Hamiltonian system p˙ j =

∂H ∂H , q˙ j = − , j = 1, 2, ... ∂qj ∂qj

where the Hamiltonian is H= 1

1 2

∞

∑ λ j (p2j + q2j ) + ε G(q)

j=1

We assume for simplicity that all eigenvalues are positive.

(2)

KAM theory with applications to Hamiltonian partial differential equations

159

and the nonlinear term is expressed in terms of the eigenfunction expansion by G(q) = ε

∑

Gi jkl qi q j qk ql , Gi jkl = 

i, j,k,l

1 λi λ j λk λl

 π 0

φi φ j φk φl dx .

Given a positive integer N and a vector ξ = (ξ1 , ..., ξN ) ∈ RN+ . Let   q j = 2(I j + ξ j ) cos θ j , p j = 2(I j + ξN ) sin θ j , j = 1, ..., N

ω = (λ1 , ..., λN ), qˆ = (q1 , ..., qN ), q˜ = (qN+1 , qN+2 , ...). Then the Hamiltonian (2) reads H = (ω , I) +

1 2

∞

∑

  λ j p2j + q2j + ε R(I, θ , q) ˆ

(3)

j=N+1

ˆ = G(q, ˜ q). ˆ Here R is independent of p˜ = (pN+1 , pN+2 , ...). More genwith R(I, θ , q) ˜ p). ˜ Let z j = (q j , p j ) erally, we can assume that R depends on p, ˜ that is, R = R(θ , I, q, and |z j |2 = |q j |2 + |p j |2 . Then (3) reads H = (ω , I) +

1 2

∞

∑

λ j |z j |2 + ε R(I, θ , z)

(4)

j=N+1

Write Λ = diag(λ j+N : j = 1, 2, ...), J = diag(J j : j = 1, 2, ...) with   0 1 . Jj = J = −1 0 The Hamiltonian vector field is ⎧ ⎪ θ˙ = ∂∂HI = ω + ε ∂∂RI ⎪ ⎪ ⎨ I˙ = − ∂∂ Hθ = −ε ∂∂ θR ⎪ ⎪ ⎪ ⎩ z˙ = J ∂∂Hz = JΛ z + ε ∂∂Rz .

(5)

We see that when ε = 0, the manifold T0 := {θ = ω t} × {I = 0} × {z = 0} is an invariant N-torus. KAM theory states that for “most” ω the invariant torus can be preserved if ε is sufficiently small. The basic idea is to seek a symplectic transformation (which is the composition of a series of transformations) with which to kill or to eliminate the perturbation R. However, up to present one has not found a symplectic transformation whish will kill the whole term R. A revised idea is to kill all lower order (
2) terms (l.o.t.) of R, that is, the linear part of vector field XR . More precisely, expanding R in a Fourier–Taylor series

160

X. Yuan

R = Rθ (θ ) + RI (θ ) · I + Rz (θ ), z + Rzz (θ )z, z + O(|I|2 + |I||z| + |z|3 )

(l.o.t.)

(h.o.t.)

If we can find a symplectic transform Ψ such that 1 H˜ = H ◦ Ψ = (ω˜ , I) + ∑ λ˜ j |z j |2 + R˜ 2 j where R˜ = O(|I|2 + |I||z| + |z|3 ), then (5) reads ⎧ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎩

θ˙ =

∂H ∂I

= ω˜ + ε ∂∂RI = ω˜ + ε O(|I|)

I˙ = − ∂∂ Hθ = −ε ∂∂ θR = −ε O(|I|2 + |I||z| + |z|3 )

(6)

z˙ = J ∂∂Hz = JΛ z + ε ∂∂Rz = JΛ z + ε O(|I| + |z|2 ).

We see that T0 is still an invariant torus of (6). Thus, Ψ −1 (T0 ) is an invariant torus of the original hamiltonian H. In searching for the symplectic transformation Ψ , one will encounter the following small divisors problems:  In order to eliminate the terms Rθ (θ ) and (RI (θ ), I), one needs conditions: (k, ω ) = 0, for all 0 = k ∈ ZN  In order to eliminate the term (Rz (θ ), z), the following Melnikov’s first conditions are required: (k, ω ) + λ j = 0, for all k ∈ ZN , j = 1, 2, ...  In order to eliminate the term (Rzz (θ )z, z), the following Melnikov’s second conditions are required: (k, ω ) + λi ± λ j = 0, for all k ∈ ZN , j = 1, 2, ..., where k = 0 if i = j and “±” takes “−” These conditions are usually not fulfilled for all ω . For example, if ω ∈ QN , then there exists k ∈ ZN such that (k, ω ) = 0. The method of addressing this problem is to regard ω as a parameter vector (or, equivalently, assume ω = ω (ξ ) depends on a parameter vector ξ and det(∂ ω /∂ ξ ) = 0). Eliminating those ω which violate the previous conditions, one can prove that the set of remaining parameters has positive measure (in a certain sense). Therefore, one has the following KAM theorem. Theorem 1 Assume λi = λ j for i = j and R is analytic in some neighborhood of the origin. Then for “most” parameters ω , there exists a symplectic transformation Ψ ˜ therefore, H possesses an invariant torus Ψ −1 (T0 ). such that H is changed into H,

KAM theory with applications to Hamiltonian partial differential equations

161

The finite dimensional version of this theorem is due to Melnikov [13, 14], Eliasson [8] and Pöschel [16]. The infinite dimensional version is due to Kuksin [10, 11], Wayne [20] and Pöschel [17]. Applying Theorem 1 to the nonlinear wave equation (1) we get Theorem 2 (Kuksin [11]) Assume the potential V = V (x, ξ ) of (1) depends on a parameter vector ξ ∈ O ⊂ RN a compact set, with Lebesgue measure 1 such that λN ) is non-degenerate, then for “most” parameters the Jacobian matrix ∂∂ ωξ = ∂ (λ1∂,..., ξ ξ (i.e., there is a subset O1 ⊂ O with Measure(O1 ) tending to zero as ε → 0 such that for ξ ∈ O \ O1 ) there is an invariant torus for (1). The motion on the torus is quasi-periodic with frequency ω˜ with |ω − ω˜ | < ε . Wayne [20] also obtains the existence of the quasi-periodic solutions of (1) when the potential V does not belong to some set of “bad” potentials. In [20], the set of all potentials is given a Gaussian measure and the set of “bad” potentials is proved to be of small measure. Because parameters are needed in Theorem 1, the potential V is assumed to depend on parameters ξ in [11] or the V itself is regarded as parameters in [20]. An important question is what happens when V does not contains any parameters. In this direction, early approaches are due to Bobenko–Kuksin [1] and Pöschel [18]. They assume V ≡ m where the constant m = 0. In [1], the term mu + u3 is regarded as a perturbation of sin u. Thus, (1) is a perturbation of sine-Gordon equation. The latter are known to be integrable, exhibiting many quasi-periodic solutions. They serves as the starting point of KAM theory in (1). An alternative method is using Birkhoff normal. Observe that for m > 0 cm , n = min{|i|, | j|, |k|, |l|} |λi ± λ j ± λk ± λl |  √ 3 n2 + m

(7)

where c is some absolute constant. This inequality allows Pöschel [18] to extract some parameters from the nonlinear term u3 through Birkhoff normal form. Once the parameters are obtained, one can apply Theorem 1 to (1). Theorem 3 ([1, 18]) For V ≡ m > 0, (1) possesses many invariant elliptic tori, and thus quasi-periodic solutions. According to Remark 7 of [18], when m ∈ (0, 1) the theorem still holds. In [21] it is shown that (1) possesses many invariant hyperbolic-elliptic tori and quasi-periodic solutions, when m ∈ (−∞, −1) \ Z. In the case V ≡ m = 0, the equation (1) is called completely resonant in [18]. In this case, One can see that the inequality (7) is useless. Whether there exists invariant torus is a challenging question, which is proposed or concerned by many authors. See references [7, 12, 15, 18]. Observe that ordinary differential y¨ + y3 = 0 is integrable and all non-zero solutions are periodic and their periods depend on amplitudes or initial values. Those solutions are also the solutions of (1), and they are uniform in the space x. Partial resonances can be overcome if we restrict ourselves to look for invariant tori at the neighborhood of those periodic solutions. Consequently, we have

162

X. Yuan

Theorem 4 ([23]) In the neighborhood of the small solutions of y¨ + y3 = 0, the equation (1) subject to periodic boundary conditions has many invariant tori of any dimension and thus quasi-periodic solutions.2 Another question is what happens when V is given but not constant, such as V = sin x, cos x. Observe that for a given potential V sufficiently smooth,

λj = j + where [V ] =

π 0

[V ] + O(1/ j2 ) j

V (x) dx. Whereas

λj =



j2 + m = j +

m + O(1/ j2 ) j

when V ≡ m = 0. By comparing these two asymptotic formulae and carefully checking the inequality (7), we have Theorem 5 ([24]) For any given potential V sufficiently smooth and [V ] = 0, the equation (1) subject to Dirichlet boundary conditions has many invariant tori of any dimension and thus quasi-periodic solutions. So far, we have a clear comprehension of the invariant tori and quasi-periodic solutions of (1). When Hamiltonian partial differential equations with spatial dimension greater than 1 are considered, a significant new problem arises due to the presence of clusters of normal frequencies of the Hamiltonian systems defined by these PDEs. For example, let us consider the higher dimensional nonlinear wave utt − $u + mu + g(x, u) = 0,

(8)

subject to Dirichlet b. c. or periodic b. c., where $ is the Laplacian in d-dimensions with d > 1. In this case, the eigenvalues λ j2 ( j ∈ Zd ) of the eigenvalues of the operator −$ + m have formula

λ j2 = | j|2 + m, j ∈ Zd . It follows that lim { j ∈ Zd : λ j = n2 + m}
= +∞

n→+∞

(9)

where
denotes the cardinality of the set. Recall the Melnikov’s second conditions (k, ω ) + λi − λ j = 0, i = j . By letting k = 0, it follows that λ j is simple, i.e., λi = λ j if i = j. Therefore, the formula (9) violates seriously Melnikov’s second conditions. In 2002, by observing some symmetries in (8), the present author [22] showed that there are many 2

There are many authors who investigate periodic solutions and 2-D quasi-periodic solutions of travelling wave type. These excellent works are less related to KAM theory. I do not present them here, because of limit of space in this talk.

KAM theory with applications to Hamiltonian partial differential equations

163

quasi-periodic solutions of traveling wave type for any spatial dimension d. Here the difficulty of small divisors was avoided owing to the symmetries. However, one can not avoid this difficulty in the search for more general solutions. In a series of papers [2] through [6], Bourgain developed another profound approach which was originally proposed by Craig–Wayne in [7], in order to overcome the difficulty that the second Melnikov’s conditions can not be imposed. Now this approach is called the C-W-B method. Instead of KAM theory, C-W-B method is based on a generalization of Lyapunov–Schmidt procedure and a technique by Fröhlich and Spencer [9]. The quasi-periodic solutions are constructed directly by Newton iteration. In that direction, one will has to investigate the inverse of a “big” matrix where the small divisors problem arises. The Fröhlich and Spencer technique is used to analyze the inverse. Usually the C-W-B method is very complicated and hard to access. Recently the present author [25] modified the classic KAM technique to avoid the second Melnikov’s conditions and succeeded to derive a new KAM theorem which can be applied to many kinds of PDEs including (8). Since the Fröhlich and Spencer technique is not needed there, the new KAM theorem is relatively easy to access. The whole of my lectures are devoted to the following KAM theorem, which appears in [25]: Theorem 6 Assume λi  c|i|κ1 and { j : λ j = λ| j| }

κ2 where c, κ1 , κ2 are absolute positive constants. Assume R is analytic in some neighborhood of the origin. Then for “most” parameters ω , there exists a symplectic transformation Ψ such that H is ˜ therefore, H possesses an invariant torus Ψ −1 (T0 ). changed into H,

2 Derivation of the linearized equations As stated in §1, the key point of KAM theory is to eliminate the (lower order) perturbation by a series of symplectic transformations which are generated by systems of linearized equations. In this section, we will derive the linearized equations. Before doing so, we introduce some notation. Let H p be the space of sequences z = (z1 , z2 ) = ((z1 j , z2 j ) ∈ C2 : j ∈ Zd ) satisfying ||z||2p =

∑ (|z1 j |2 + |z2 j |2 )| j|2p < ∞

j∈Zd

where d is the dimension of the Laplacian and p > d/2 is given. It is easy to see that H p is a Hilbert space with an inner product corresponding to the norm || · || p . (In fact, H p corresponds to the so-called Sobolev space H p by means of Fourier transform.) Denote by L(H p , H p ) all bounded linear operators from H p to H p . Introduce the phase space: P := (Cn /2π Zn ) × Cn × H p , where n is a given positive integer. We endow P with a symplectic structure √ √ d θ ∧ dy + −1dz1 ∧ dz2 = dx ∧ dy + −1 ∑ dz1 j ∧ dz2 j , (θ , I, z1 , z2 ) ∈ P. j∈Zd

164

X. Yuan

Given r, s > 0. Define a domain in P by D(s, r) = {(θ , I, z) ∈ P : | Im θ |
s, |I|
r2 , ||z|| p
r} and D(s) = {θ ∈ Cn /2π Zn : | Im θ |
s} . For z, z˜ ∈ H p , define z, z˜ := ∑(z1 j z˜1 j + z2 j z˜2 j ). j

For a sequence of real numbers {λ j : j ∈ Zd }, let     10 d : j∈Z . Λ = diag λ j 01 Note ·, · is not an inner product. Consider a Hamiltonian H defined on D(s, r): H = N +R where

1 1 N = (ω , I) + Λ z, z + B(θ )z, z (1) 2 2 and R = R(θ , I, z) : D(s, r) → C and B = B(θ ) : D(s) → L(H p , H p ) are analytic. As in §1, write 1 R = Rθ (θ ) + RI (θ ) · I + Rz (θ ), z + Rzz (θ )z, z 2 2

+ O(|I|

+ |I|||z|| p + ||z||3p )

(l.o.t.) (2)

(h.o.t.)

The basic idea is to kill the lower terms (l.o.t.). As stated in §1, in order to eliminate the term Rzz (θ )z, z we need to assume the second Melnikov conditions, which prevents the KAM theorem from being applied to higher dimensional PDEs. Hence we should modify the basic idea. Following Bourgain [6], we put the term Rzz (θ )z, z into the “integrable” part N rather than to eliminate it. However, doing so will make the problem too complicated. We just want to put a part of Rzz (θ )z, z into N. To this end, we introduce a cut-off operator Γ as follows. Given a positive number K large enough. For any an operator or (vector) function f defined on the domain D(s), write √ f = ∑ f-(k)e −1(k,θ ) . k∈Zn

Define the cut-off operator: (Γ f )(θ ) = (ΓK )(θ ) :=

∑

|k|
K

√ −1(k,θ )

f-(k)e

.

KAM theory with applications to Hamiltonian partial differential equations

165

The modified KAM procedure consists of the following steps. Step 1. Averaging and cut-off. Recall (1) and (2). H =N +R 1 = (ω , I) + Λ z, z + B(θ )z, z 2 1 + Rθ (θ ) + RI (θ ) · I + Rz (θ ), z + Rzz (θ )z, z (l.o.t.) 2 + O(|I|2 + |I|||z|| p + ||z||3p ) 1 = (ω + R-I (0), I) + Λ z, z + (B(θ ) + Γ Rzz (θ ))z, z (:= N) 2 + Γ Rθ + (Γ RI − R-I (0)) · I + Γ Rz , z (:= R1 ) 1 + (1 − Γ )Rθ + (1 − Γ )RI · I + (1 − Γ )Rz , z + (1 − Γ )Rzz z, z (:= R2 ) 2 + O(|I|2 + |I|||z|| p + ||z||3p ) (:= R3 ) = N + R1 + R2 + R3

(3)

Assume Γ B(θ ) = B(θ ). Let ω = ω + R-I (0) and B(θ ) = B(θ ) + Γ Rzz (θ ). Then 1 1 N = (ω , I) + Λ z, z + B(θ )z, z. 2 2 .θ (0) = 0, since any Notice that Γ B = B, Γ R1 = R1 . By the way, we can assume R constant added to the Hamiltonian function does not affect the dynamics. Step 2. Seek a symplectic transformation to eliminate the term R1 . Assume we have a Hamiltonian function of the same form as R1 : F = F θ (θ ) + F I (θ ) · I + F z (θ ), z .θ (0) = 0, F .I (0) = 0 and Γ F = F. Denote by X t the flow of the vector field where F F XF corresponding to the Hamiltonian function F. Let Ψ = XF1 , it is a symplectic transformation. Let (θ (t), I(t), z(t)) be a solution of the vector field XF , that is,

∂F ˙ ∂F , I=− θ˙ = , z˙ = J ∂z F. ∂I ∂θ (Referring to (5).) Then we have d d H ◦ XFt = H(θ (t), I(t), z(t)) dt dt ∂H ˙ ∂H ˙ ∂H z˙ θ+ = I+ ∂θ ∂I ∂z   ∂H ∂F ∂H ∂F ∂H ∂F = − J (θ (t), I(t), z(t)) + ∂θ ∂I ∂I ∂θ ∂z ∂z := {H, F} ◦ XFt

(4)

166

X. Yuan

By abuse of notation, we still denote by (θ , I, z) the new variables XF1 (θ , I, z). By Taylor’s formula and (4), we get H ◦ Ψ = H ◦ XF1 = H ◦ XF0 +

d 1 H ◦ XFt |t=0 + dt 2

 1 0

 1 1

(1 − t)

d2 H ◦ XFt dt dt 2

(1 − t){{H, F}, F} ◦ XFt 2 0 = N + R1 + {N, F} + R˜ = H + {H, F} +

1 1 = (ω , I) + Λ z, z + B(θ )z, z + Γ Rθ + (Γ RI − R-I (0)) · I + Γ Rz , z 2 2 + {ω · ∂θ F θ + (ω · ∂θ F I ) · I + ω · ∂θ F z + Λ JF z + Γ (BJF z ), z + (1 − Γ )(BJF z ), z + (∂θ B, F I )z, z} + R˜ where 1 R˜ = R2 + R3 + {R1 + R2 + R3 , F} + 2

 1 0

(1 − t){{H, F}, F} ◦ XFt .

If we can find F solving the following linear equations:

ω · ∂θ F θ = Γ Rθ ,

(5)

ω · ∂θ F I = Γ RI − R-I (0),

(6)

ω · ∂θ F z + Λ JF z + Γ (BJF z ) = Γ Rz ,

(7)

and write B+ = B + (∂θ B, F I ), 1 1 N+ = (ω , I) + Λ z, z + B+ (θ )z, z, 2 2 z ˜ R+ = R + (1 − Γ )(BJF ), z , then we get H+ = H ◦ Ψ = N+ + R+ .

(8)

In §3, we will show that if R = O(ε ) in some domain, then F = O(ε 1− ) in a smaller sub-domain which is the result of excision of some parameters ω of small total measure. Therefore, R+ = O(ε 4/3 ). Repeating the procedure above m-times, then m R+ = O(ε (4/3) ). Let m → +∞. Then we get a Hamiltonian 1 H∞ = N∞ = (ω˜ , I) + (λ + B(θ ))z, z . 2 This completes the proof of Theorem 6.

KAM theory with applications to Hamiltonian partial differential equations

167

3 Solutions of the linearized equations We are now in position to solve the homological equations (5, 6, 7). Solving (5, 6) is easy. We focus our attention on the solution of (7). Observe that  √      √ 1 1 1 0 0 1 √−1 Q0 , Q0 = √ = −1Q∗0 0 −1 −1 0 2 1 − −1 where ∗ is the conjugate transpose of the matrix. It is easy to see that Q∗0 = Q−1 0 . Let    
1 0 : j ∈ Zd . Q = diag Q0 : j ∈ Zd , E± = diag 0 −1 √ Then J = −1Q∗ E± Q and QΛ Q∗ = Λ . Notice that E02 is the identity operator of 2 (Zd ) ⊗ 2 (Zd ). Left-multiplying (7) by Q we get √ √ ω E± ∂θ (E± QF z ) + −1QΛ Q∗ (E± QF z ) + −1Γ (QBQ∗ (E± QF z )) = Γ QRz . Let Fz = E± QF z , Rz = QRz and B = QBQ∗ . Note Γ Q = QΓ . Then √ √ ω E± ∂θ Fz + −1Λ Fz + −1Γ (BFz ) = Γ Rz . Write Fz =

∑

√ −1(k,θ )

F(k)e

|k|
K

Rz =

√

−1

, B=

∑

∑

|k|
K √ −1(k,θ )

R(k)e

√ −1(k,θ )

B(k)e

(1)

,

.

|k|
K

Let T = diag(±(k, ω ) + λ j : |k|
K, j ∈ Zd , k ∈ Zn ) - = (B(k - − l) : |k|, |l|
K, k, l ∈ Zn ) B - : |k|,
K, k ∈ Zn ), R- = (R(k) - : |k|,
K, k ∈ Zn ). F- = (F(k) Then (7) can be written as

- F- = R. (T + B)

(2)

- is invertible and to find its inverse. Our goal is to prove that the operator T + B To this end, we need some assumptions. A1. Assume ω = ω (ξ ), B = B(θ , ξ ) depend smoothly (in the sense of Whitney3 ) on a parameter vector ξ ∈ O a compact set with Meas(O) > 0, and

sup θ ∈D(s),ξ ∈O 3

|det(∂ ω /∂ ξ )|  C > 0 ,   ||B(θ , ξ )|| p  1, sup ∂ξ B(θ , ξ ) p  1.

We will not mention this further.

θ ∈D(s),ξ ∈O

(3) (4)

168

X. Yuan

We denote by C a universal positive constant whose value may be different in differ
ent places. Let K
= Cardinality of {k ∈ Zn : |k|
K} and H p = H p ⊗ CK with norm ||u||2p =

K


∑ ||u j ||2p , ∀ u = (u1 , ..., uK
) ∈ H p .

j=1

- is bounded linear operator in H p , and It follows from (8) that B - p  1, ||∂ξ B|| - p  1, ∀ξ ∈ O ||B||

(5)

where || · || p is the operator norm of H p . A2. Assume the original Hamiltonian H = N + R is real for real arguments. It follows that for (θ , ξ ) ∈ D(s) × O with s = 0, the operator B(θ , ξ ) is self-adjoint in the space 2 (Zd ) × 2 (Zd ) (note 2 (Zd ) × 2 (Zd ) = H p with p = 0). If we regard T

B as a matrix of infinite dimension, then B(θ , ξ ) = B(θ , ξ ) where the bar means conjugate and T means transpose. Recall that B = QBQ¯ T . We have T

B(θ , ξ ) = B(θ , ξ ), ∀ (θ , ξ ) ∈ D(0) × O. It follows from this that T

- − k), ∀ k, l ∈ Zn . - − l) = B(l B(k

(6)

- is self-adjoint in the space 2 := 2 (Zd ) × 2 (Zd ) × Then it follows from (6) that B
- is of infinite dimension. We will reduce the inverse of CK . Note the matrix T + B it to one of a matrix of finite dimension. To this end, we need a third assumption. A3. Assume the normal frequencies λ j ’s satisfy the following growth conditions:

λ j  C| j|κ , ∃ κ > 0, ∀ j ∈ Zd . Let

 M = 1+

supξ |ω (ξ )| K C

1/κ

:= CK 1/κ .

We see that when |k|
K, | j|  M, | ± (k, ω ) + λ j |  1 . Write T1 = diag(±(k, ω ) + λ j : |k|
K, | j|
M) , T2 = diag(±(k, ω ) + λ j : |k|
K, | j|  M) .

(7)

KAM theory with applications to Hamiltonian partial differential equations

169

Then T = T1 ⊕ T2 . And by (7) we get ||T2−1 || p
1. - as a matrix of infinite dimension, we denote by B - i j (k − l) the While regarding B d n - into four elements of matrix where i, j ∈ Z , k, l ∈ Z , |k|, |l|
K. We decompose B blocks as follows: - 11 = (Bi j (k − l) : |k|, |l|
K, |i|, | j|
M) B - 12 = (Bi j (k − l) : |k|, |l|
K, |i|
M, | j| > M) B - 21 = (Bi j (k − l) : |k|, |l|
K, |i| > M, | j|
M) B - 22 = (Bi j (k − l) : |k|, |l|
K, |i| > M, | j| > M). B 

Then -= B

- 11 B - 12 B - 21 B - 22 B

 .

According to (5), - i j || p  1, ||∂ξ B - i j || p  1, ∀ξ ∈ O, i, j ∈ {1, 2}. ||B

(8)

- 22 || p  1, we get By ||T2−1 || p
C and ||B ∞

- 22 )−1 || p
|| ∑ ||(T −1 B - 22 ) j || p ||T −1 ||| p
C. ||(T2 + B 2 2 j=0

Set

 0 E1 −(T2 + B-22 )−1 B-21 E2   E1 −B-12 (T2 + B-22 )−1 , Brig := 0 E2 B-le f :=



where E1 (E2 , respectively) is a unit matrix of the same order as that of T1 (T2 , - 22 )−1 || p
C, it is easy to verify that respectively). From ||(T2 + B - −1 || p
C - −1 || p , |||B ||B rig le f Let

- 11 − B - 12 (T2 + B - 22 )−1 B - 21 -11 = B B

- exists; Then the inverse of T + B   −1 + B ) 0 (T 1 11 - −1 - −1 - −1 = B (T + B) rig - 22 )−1 Ble f 0 (T2 + B

170

X. Yuan

and

- −1 || p
C||(T1 + B -11 )−1 || p , ||(T + B)

-11 exists. provided that the inverse of T1 + B -11 . First of all, we We are now in position to investigate the inverse of T1 + B would like to point out that the matrix B11 is of finite order, and the order is bounded by −1 K ∗ := K 2n M d
K 2n+d κ . - that the matrix B -11 is also selfSecondly, it follows from the self-adjointness of B adjoint. Thirdly, by (8) we have -11 || p  1 . -11 || p  1, qquad||∂ξ B ||B - is continuously differentiable in ξ ∈ O, the matrix Fourthly, since each element of B B11 is also continuously differentiable in ξ ∈ O. In view of (3), we can regard ω itself as a parameter vector instead of ξ , or we can assume ω = ξ . Without loss of generality, we assume the first entry ω1 of ω is in the interval [1, 2]. Then the matrix -11 is non-singular if and only if T1 + B -11 A1 := ω1−1 T1 + ω1−1 B is non-singular, and

-11 )−1 || p
C||A−1 || p . ||(T1 + B 1

Let

Ξ : ς1 = 1/ω1 , ς2 = ω2 /ω1 , ..., ςn = ωn /ω1 . Then it is easy to get     det ∂ (ς1 , ..., ςn )  = ω −(n+1)  C > 0 . 1  ∂ (ω1 , ..., ωn ) 

Therefore, we can regard ς as a parameter vector. It is easy to see that Meas O
C Meas Ξ (O)
C Meas O and

-11 (ξ (ς ))||  1 , ||B

-11 (ξ (ς ))||  1 . ||∂ς B

After introducing the parameter ς , we can write n

A1 = diag (±(k1 + ∑ kl ςl ) + ς1 Ω 0j : k = (k1 , ..., kn ) ∈ Zn , |k|
K, | j|
M) l=2

-11 (ξ (ς )). + ς1 B

(9)

KAM theory with applications to Hamiltonian partial differential equations

171

-11 is self-adjoint, so is A1 = A1 (ς ) for any ς ∈ Ξ (O). Therefore, there are Since B continuously differentiable functions µ1 (ς ), · · · , µK ∗ (ς ) representing the eigenvalues of A1 for ς ∈ Ξ (O). Lemma 2 There exists a subset O+ ⊂ O with Meas(Ξ O+ )  (MeasΞ (O))(1 − CK −1 such that for any ς ∈ Ξ (O+ ),

µ j (ς )  (KK ∗ )−1 > 0. We postpone the proof to the end of this section. Since A1 is self-adjoint, there exists a matrix-valued function U(ς ) of order K ∗ which depends on ς , such that for every ς ∈ Ξ (O+ ) the following equalities hold: A1 (ς ) = U(ς )diag(µ1 (ς ), · · · , µK ∗ (ς ))U ∗ (ς ), and

U(ς )(U(ς ))∗ = (U(ς ))∗U(ς ) = E

where E is the unit matrix of order K ∗ and U ∗ is the conjugate transpose of U. It follows that for ς ∈ Ξ (O+ ), ||A1 (ς )||
max{µ j : j = 1, ..., K ∗ }
KK ∗ where || · || is the 2 norm of matrix. Since A1 is of order K ∗ , ||A1 (ς )|| p
KK ∗ (K ∗ ) p := KC . Thus, - −1 || p
||(T1 + B -11 )−1 || p
||A1 (ς )|| p
KC , ξ ∈ O+ . ||(T + B) Assume that sup ||∂ξl Rz (θ , ξ ))|| p
ε , l = 0, 1 .

D(s)×O

It follows that - p
KC sup ||Rz (θ , ξ ))|| p
KC ε . - p
KC ||R|| ||F|| D(s)×O

-11 )||| p
K. We have that for any ξ ∈ O+ , Note |||∂ξ (T1 + B -11 )−1 || p =||(T1 +B -11 )−1 (∂ξ (T1 + B -11))(T1 + B -11 )−1 || p
K 1+2C := KC . ||∂ξ (T1 +B It follows that Moreover,

- −1 || p
KC , ξ ∈ O+ . ||∂ξ (T + B) - p
KC ε . ||∂ξ F||

172

X. Yuan

Note

- 2p = ||F||

∑

|k|
K

.z (k)||2p + ||F

∑

.z¯ (k)||2 . ||F p

∑

.u (k)e F

|k|
K

Then sup

||F u (x, ξ )||2p =

D(0)×O+

||

sup D(0)×O+

∑




|k|
K

|k|
K

√ −1(k,x) 2 || p

.u (k)||2p
(KC ε )2 ||F

that is, ||F u (x, ξ )|| p
KC ε .

sup

(10)

D(0)×O+

Similarly, sup D(0)×O+

||∂ξ F u (x, ξ )|| p
KC ε .

Lemma 3 We can extend the domain D(0) to D(s) such the above inequalities still hold; sup ||F u (x, ξ )|| p
KC ε , D(s)×O+

sup ||∂ξ F u (x, ξ )|| p
KC ε .

D(s)×O+

Proof. Rewrite F u = (F z , F z¯ )) and  B=

 Bzz Bz¯z . Bz¯z Bz¯z¯

Then the homological equation (7) can be rewritten as √ √ − −1ω · ∂θ F z + Ω 0 F z + Γ ((Γ Bz¯z )F z − (Γ Bzz )F z¯ ) = − −1Γ Rz (x, ξ ) , √ √ −1ω · ∂θ F z¯ + Ω 0 F z¯ + Γ ((Γ Bz¯z )F z¯ − (Γ Bz¯z¯ )F z ) = −1Γ Rz¯ (x, ξ ) . The following equalities can be fulfilled by the assumption that H is real for real argument: ¯ Rz = Rz¯ , Bzz = Bz¯z¯ , Bz¯z = Buu = Bzz = Buu¯ , θ ∈ D(0).

See [25] for the details. It follows that F z (θ ) = F z¯ (θ ) for θ ∈ D(0). Note F z is analytic in D(s). Thus, F z (θ ) = F z¯ (θ ) for θ ∈ D(s). Let i2 = −1. In the proof of this lemma, we can assume ω = (1, ..., 1) without loss of generality. An important fact is that λ j ’s are positive. When the dimension n = 1 of the angle variable θ , the proof shows more clearly our basic idea. Firstly, assume n = 1. Let θ = it + r. Arbitrarily fix r ∈ R/2π Z. Write F(t) = F z¯ (it + r). By the second homological equation and using the method of variation of constants,

KAM theory with applications to Hamiltonian partial differential equations

F(t) = F z¯ (r) −

 t 0

e−Λ (t−τ ) (BF + Rz (it)))d τ , t ∈ [0, s] .

173

(11)

Let || · || p,s = supD(s) || · || p where we write formally BF = Γ [Γ Bz¯z (iτ ))F(iτ ) + (Γ Bz¯z )F(it)]. Note ||B|| p,s
δ  1, ||Rz || p,s
ε . And note ||F z || p = ||F z¯ || p . By (10) we have ||F z¯ (r)|| p = ||F z (r)|| p
KC ε . Therefore, by (11) we get ||F(t)|| p
KC ε + δ ||F z || p,s , t ∈ [0, s]. By the first homological equation, we get ||F(it)|| p
KC ε + δ ||F z || p,s , t ∈ [−s, 0]. Thus, ||F(it)|| p
KC ε + δ ||F z || p,s , t ∈ [−s, s]. That means ||F z¯ (it + r)|| p
KC ε + δ ||F z || p,s , t ∈ [−s, s], r ∈ R/2π Z. This leads to ||F z || p,s
2KC ε := KC ε . That is sup ||F z (x, ξ )|| p
KC ε . D(s)×O+

Now let us consider the dimension n = 2. Fix an arbitrary r ∈ R/2π Z. Let θ1 = it +r with t ∈ [0, s], let F(t) := F z¯ (it +r, φ ), and restrict φ ∈ R. By the second homological equation, we have F(t) = F z¯ (r, φ ) −

 t 0

e−Λ (t−τ ) e−i(t−τ )∂φ (BF + Rz ), t ∈ [0, s]

(12)

For any analytic 2π -periodic function f : {x : | Im x|
s} → H p , using Cauchy’s theorem, we have sup ||e−i(t−τ )∂x f (x)|| p
e|t−τ |/s || f || p,s .

x∈R/2π Z

By (12) we have ||F z¯ (it + r, φ )|| p
||F z¯ (r, φ )|| p +

 t 0


KC ε + δ ||F z¯ || p,s ,

e(t−τ )/s [δ ||F z¯ || p,s + ||Rz || p,s ]d τ t ∈ [0, s], φ ∈ R.

174

X. Yuan

Again by the first homological equation and noting F z = F z¯ , we have ||F z¯ (it + r, φ )|| p
KC ε + δ ||F z¯ || p,s ,

t ∈ [−s, s], φ ∈ R

(13)

For any constant c, the line L = L(c) : x − y = c is a characteristic line of ∂x + ∂y . Let F(y) := F(i(y + c) + r1 , iy + r2 ), y ∈ [0, s], r1 , r2 fixed. By the second homological equation, we have F(y) = F z¯ (ic + r1 , r2 ) −

 y 0

e−Λ (y−τ ) (BF + Rz )d τ , y ∈ [0, s].

(14)

It follows that ||F z¯ (i(y + c) + r1 , iy + r2 )|| p
||F z¯ (ic + r1 , r2 )|| p + δ ||F z¯ || p,s + ε , y ∈ [0, s]. Moreover, by the first homological equation, we have ||F z¯ (i(y + c) + r1 , iy + r2 )|| p
||F z¯ (ic + r1 , r2 )|| p + δ ||F z¯ || p,s + ε , y + c, y ∈ [−s, s]. By (13), ||F z¯ (i(y + c) + r1 , iy + r2 )|| p
KC ε + 2δ ||F z¯ || p,s + ε , y + c, y ∈ [−s, s]. Let the line L(c) run over the square [−s, s]2 , we have ||F z¯ || p,s
KC ε + 2δ ||F z¯ || p,s + ε . It follows ||F z¯ || p,s
4KC ε := KC ε . We will omit the proof of ||∂ξ F z¯ || p,s
KC ε . This proof is finished by mathematical induction on n. Proof of Lemma 2. Let

Ξ (Ol ) = {ς ∈ Ξ (O) : |µl | < 1/(KK ∗ )}, l = 1, ..., K ∗ . Take an arbitrary µ = µ (ς ) ∈ {µ1 (ς ), ..., µK ∗ (ς )}. Let φ be the normalized eigenvector corresponding µ . It is easy to prove that ∂ς µ = ((∂ς A1 )φ , φ ). By computing ∂ς A1 and using (9) and Lemma 4, we get

∂ς µ = ((diag(λ j : |k|
K, | j|
M))φ , φ ) + o(1)  min{λ j : j ∈ Zd } + o(1)  C > 0. It follows that Meas Ξ (Ol )
1/(KK ∗ ). Thus, ∗

Meas

K / l=1

Ξ (Ol ) < 1/K.

KAM theory with applications to Hamiltonian partial differential equations

Let

175

∗

Ξ (O) = Ξ (O) \

K /

Ξ (Ol ).

l=1

Therefore,

1 Meas Ξ (O)  Meas Ξ (O)(1 − O( )) K

and for any ς ∈ Ξ (O),

|µl (ς )|  1/KK ∗ .

This completes the proof of Lemma 2.

4 Applications to partial differential equations in higher dimensions We will just give the application of Theorem 6 to a nonlinear wave equation in higher dimension. See [25] for other applications such as to the nonlinear Schrödinger equation. Consider the nonlinear wave equation utt − $u + Mσ u + ε u3 = 0, θ ∈ Td , d  1

(1)

where u = u(t, θ ) and $ = ∑dj=1 ∂θ2j and Mσ is a real Fourier multiplier Mσ cos( j, θ ) = σ j cos( j, θ ), Mσ sin( j, θ ) = σ j sin( j, θ ), σ j ∈ R, j ∈ Zd . Pick a set ϖ = {ϖ1 , ..., ϖn } ⊂ Zd . Let Zd = Zd \ {ϖ1 , ..., ϖn }. Following Bourgain [4], we assume & σϖl = σl , (l = 1, ..., n) σ j = 0, j ∈ Zd .   Theorem 7 ([25]) Let ωl0 = λϖl = |ϖl |2 + σl , (l = 1, ..., n) and σ = (σ1 , ..., σn ) and ω 0 = (ω10 , ..., ωn0 ). Then there is a subset O 0 ⊂ [1, 2]d with Meas O0  (1 − Cε ) such that for any σ ∈ O 0 , the nonlinear wave equation with small ε has a rotational invariant torus of frequency vector ω with |ω − ω 0 | = O(ε ). The motion on the torus can be expressed by u(t, θ ) which is quasi-periodic (in time) with frequency ω and u(·, θ ) : R → H p (Tn ) is an analytic map, and thus the solution u(t, θ ) is, at least, a sufficiently smooth function of (t, θ ) if p is taken large enough. Acknowledgements The author is very grateful to Prof. W. Craig for his help.

176

X. Yuan

Appendix Lemma 4 Assume the matrix A = A(ς ) is self-adjoint and smooth in ς . Let µ = µ (ς ) be any eigenvalue of A and φ be the eigenfunction corresponding to µ . Then we have ∂ς µ = ((∂ς A)φ , φ ). Proof. Note φ is not necessarily smooth in ς . Consider the difference operator: $ f = $ς1 ς2 f =

f (ς1 ) − f (ς2 ) . ς 1 − ς2

Apply $ to Aφ = µφ , ($A)φ + A($φ ) = ($µ )φ + µ ($φ ). Taking inner product with φ , ($A)φ , φ  + A($φ ), φ  = ($µ )φ , φ  + µ $φ , φ . Since A is self-adjoint, A($φ ), φ  = ($φ ), Aφ  = µ $φ , φ . Thus ($A)φ , φ  = ($µ )φ , φ  = ($µ )φ , φ  = $µ . Letting ς1 → ς1 := ς , we have ∂ς µ = ((∂ς A)φ , φ ).

References 1. Bobenko, A.I. and Kuksin, S.B., The nonlinear Klein-Gordon equation on an interval as a perturbed sine-Gordon equation, Comment. Math. Helv. 70 (1995), no. 1, 63–112. 2. Bourgain, J., Construction of quasi-periodic solutions for Hamiltonian perturbations of linear equations and application to nonlinear PDE. Int. Math. Research Notices 11 (1994), 475–497. 3. Bourgain, J., Periodic solutions of nonlinear wave equations. Harmonic analysis and partial equations, Chicago Univ. Press (1999), pp. 69–97. 4. Bourgain, J., Quasi-periodic solutions of Hamiltonian perturbations for 2D linear Schrödinger equation. Ann. Math. 148 (1998), 363–439. 5. Bourgain, J., Green’s function estimates for lattice Schrödinger operators and applications. Annals of Mathematics Studies 158 Princeton University Press, Princeton, NJ, (2005). 6. Bourgain, J., On Melnikov’s persistence problem, Math. Res. Lett. 4 (1997), 445–458. 7. Craig, W. and Wayne, C.E., Newton’s method and periodic solutions of nonlinear wave equation, Commun. Pure. Appl. Math. 46 (1993), 1409–1501. 8. Eliasson, L.H., Perturbations of stable invariant tori for Hamiltonian systems, Ann. Scula Norm. Sup. Pisa CL Sci. 15 (1988), 115–147. 9. Frohlich, J. and Spencer, T., Absence of diffusion in the Anderson tight binding model for large disorder or lower energy. Commun. Math. Phys. 88 (1983), 151–184.

KAM theory with applications to Hamiltonian partial differential equations

177

10. Kuksin, S.B., Hamiltonian perturbations of infinite-dimensional linear systems with an imaginary spectrum, Funct. Anal. Appl., 21 (1987), 192–205. 11. Kuksin, S.B., Nearly Integrable Infinite-Dimensional Hamiltonian Systems, Lecture Notes in Mathematics, vol. 1556, Springer, Berlin (1993). 12. Kuksin, S.B., Elements of a qualitative theory of Hamiltonian PDEs, Proceedings of ICM 1998, Vol. II, Doc. Math. J. DMV (1998), 819–829. 13. Melnikov, V.K., On some cases of conservation of conditionally periodic motions under a small change of the Hamiltonian function, Dokl. Akad. Nauk SSSR 165:6 (1965), 1245–1248; Sov. Math. Dokl. 6 (1965), 1592–1596. 14. Melnikov, V.K., A family of conditionally periodic solutions of a Hamiltonian system, Dokl. Akad. Nauk SSSR 181:3 (1968), 546–549; Sov. Math. Dokl. 9 (1968), 882–886. 15. Marmi, S. and Yoccoz, J.-C., Some open problems related to small divisors, (Lecture Notes in Math. 1784). Springer, New York (2002). 16. Pöschel, J., On elliptic lower dimensional tori in hamiltonian systems, Math. Z. 202 (1989), 559–608. 17. Pöschel, J., A KAM-Theorem for some nonlinear partial differential equations. Ann. Sc. Norm. Sup. Pisa 23 (1996), 119–148. 18. Pöschel, J., Quasi-periodic solution for nonlinear wave equation, Commun. Math. Helvetici 71 (1996), 269–296. 19. Rabinowitz, P.H., Free vibrations for a semilinear wave equation. Commm. Pure Appl. Math. 31 (1978), 31–68. 20. Wayne, C.E., Periodic and quasi-periodic solutions of nonlinear wave equations via KAM theory, Comm. Math. Phys, 127 (1990), no. 3, 479–528. 21. Yuan, X., Invariant manifold of hyperbolic-elliptic type for nonlinear wave equation. IMMS, 18 (2003), 1111–1136. 22. Yuan, X., Quasi-periodic solutions of nonlinear Schrödinger equations of higher dimension, J. Diff. Eq., 195 (2003), 230–242. 23. Yuan, X., Quasi-periodic solutions of completely resonant nonlinear wave equations, J. Differential Equations, 230 (2006), 213–274. 24. Yuan, X., Invariant tori of nonlinear wave equations with a prescribed potential. Discrete Continuous Dynamical Sys., 16 (2006), no. 3 615–634. 25. Yuan, X., A KAM theorem with applications to partial differential equations of higher dimensions, Commun. Math. Phys., 275 (2007), 97–137.

Four lectures on KAM for the non-linear Schrödinger equation L.H. Eliasson1 and S.B. Kuksin2

Abstract We discuss the KAM-theory for lower-dimensional tori for the non-linear Schrödinger equation with periodic boundary conditions and a convolution potential in dimension d. Central in this theory is the homological equation and a condition on the small divisors often known as the second Melnikov condition. The difficulties related to this condition are substantial when d  2. We discuss this difficulty, and we show that a block decomposition and a TöplitzLipschitz-property, present for non-linear Schrödinger equation, permit to overcome this difficuly. A detailed proof is given in [EK06].

1 The non-linear Schrödinger equation We formulate the equation as an ∞-dimensional Hamiltonian system and as a problem of persistency of lower-dimensional invariant tori.

1.1 The non-linear Schrödinger equation We consider the ∆ -dimensional nonlinear Schrödinger equation −iu˙ = −∆ u +V (x) ∗ u + ε

∂F (x, u, u) ¯ , ∂ u¯

for u = u(t, x) under the periodic boundary condition x ∈ Td . The convolution potential V : Td → C have real Fourier coefficients Vˆ (a), a ∈ Zd , and we shall 1

University of Paris 7, Department of Mathematics, Paris e-mail: [email protected]

2

Heriot-Watt University, Department of Mathematics, Edinburgh e-mail: [email protected]

W. Craig (ed.), Hamiltonian Dynamical Systems and Applications, 179–212. c 2008 Springer Science + Business Media B.V.


179

180

L.H. Eliasson and S.B. Kuksin

suppose it is analytic. (This equation is a popular model for the ‘real’ NLS equation, where instead of the convolution term V ∗ u we have the potential term Vu.) F is an analytic function in Re u, Im u and x. When F(x, u, u) ¯ = (uu) ¯ 2 this is the cubic Schrödinger equation. For ε = 0 the equation is linear and has time–quasi-periodic solutions i(|a| +V (a))t i ˆ e , ∑ u(a)e 2

u(t, x) =

ˆ

a∈A

where A is any finite subset of Zd and |u(a)| ˆ > 0. We shall treat ωa = |a|2 + ˆ V (a), a ∈ A as free parameters in some domain U ⊂ RA . For ε = 0 we have under general conditions: If |ε | is sufficiently small, then there is a large subset U of U such that for all ω ∈ U the solution u persists as a time–quasi-periodic solution which has all Lyapounov exponents equal to zero and whose linearized equation is reducible to constant coefficients. In these lectures we shall describe the basic difficulty related to this result – often known as the second Melnikov condition – and the ideas behind its solution. A detailed proof is given in [EK06].

1.2 An ∞ -dimensional Hamiltonian system We write

&

u(x) = ∑a∈Zd ua ei u(x) = ∑a∈Zd va ei Zd

In the symplectic space {(ua , va ) : a ∈ Zd } = C i

∑

(va = u¯a ). × CZ , d

dua ∧ dva ,

a∈Zd

the equation becomes a Hamiltonian system & u˙a = i ∂∂va (h + ε f )

v˙a = −i ∂∂ua (h + ε f )

a ∈ Zd ,

with an integrable part h(u, v) =

∑ (|a|2 + Vˆ (a))ua va

a∈Zd

plus a perturbation

ε f (u, v) = ε

1 (2π )d

 Td

F(x, u(x), u(x))dx.

Four lectures on KAM for the non-linear Schrödinger equation

181

The second derivatives of f have a Töplitz invariance:

∂2 f ∂2 f = ∂ ua+c ∂ vb+c ∂ ua ∂ vb and

∂2 f ∂2 f = ∂ ua+c ∂ ub−c ∂ ua ∂ ub

(and similar for the second derivatives with respect to va , vb ), for any c ∈ Zd . This is easy to see for the cubic Schrödinger where f (u, v) =

∑

ua ub vc vd .

(1)

a+b−c−d=0

For example

∂2 f = 2 ∑ vc vd ∂ ua ∂ ub c+d=a+b

which clearly have this invariance. The non-linear Schrödinger is a real Hamiltonian system. Indeed if we let     ua ξa =C , ζa = ηa va with

1 C= √ 2



 1 1 , −i i

(2)

then, in the symplectic space {(ξa , ηa ) =: a ∈ Zd } = RZ × RZ , d

∑

d

d ξa ∧ d ηa ,

a∈Zd

the equation becomes &˙ ξa = − ∂ ∂ηa (h + ε f )

η˙ a =

∂ ∂ ξa (h + ε

f)

a ∈ Zd ,

also written ζ˙a = J ζ∂a (h + ε f ), with the integrable part h(ξ , η ) =

1 ∑ (|a|2 + Vˆ (a))(ξa2 + ηa2 ) 2 a∈Z d

plus the perturbation ε f (ξ , η ) which is real, because F is a real function of Re u and Im u. The Töplitz-invariance of the second derivatives can of course be formulated in these coordinates but the description is more complicated (see Sect. 5.2).

182

L.H. Eliasson and S.B. Kuksin

1.3 The topology Let L be an infinite subset of Zd . The space lγ2 (L , R),

γ  0,

is the set of sequences of real numbers ξ = {ξa : a ∈ L }, such that  ξ γ = ∑ |ξa |2 a2m∗ e2γ |a| < i a = max(|a|, 1). a∈L

There is a natural identification of lγ2 (L , R) × lγ2 (L , R), whose elements are (ξ , η ), with lγ2 (L , R2 ), whose elements are {ζa = (ξa , ηa ) : a ∈ L }, and we will not distinguish between them. We shall assume that m∗ > d2 . Then, in the phase space l02 (Zd , R2 ), our Hamiltonian h + ε f is analytic (in some domain O in l02 (Zd , R)). To see that f is analytic, consider for example the cubic Schrödinger in the complex variables (1). Using the estimate  ∑ |ua |
∑a−2m∗ u0 , a

we have

a

| f (u, v)|
u20 v20 ,

and it follows easily that f is analytic. Let denote the “pairing” =

∑



 ξa ξa + ηa ηa .

a∈Zd

Since the phase space is a Hilbert space, its first differential l02 (L , R2 )  ζˆ → defines a vector ∂ζ f (ζ ), its “gradient”, (with respect to the pairing), and its second differential 1 l02 (L , R2 )  ζˆ → 2 defines a matrix ∂ζ2 f (ζ ) : L × L → gl(2, R), its “Hessian”, (with respect to the pairing), which is symmetric, i.e. t

(

∂2 f ∂2 f (ζ )) = (ζ ). ∂ ζa ∂ ζb ∂ ζb ∂ ζa

For ζ ∈ O ∩ lγ2 (Zd , R2 ), γ > 0, the gradient and the Hessian verifies certain properties of exponential decay. These properties are most easily seen in the complex variables u, v. For example for the cubic Schrödinger (1) the gradient of f verifies

Four lectures on KAM for the non-linear Schrödinger equation

|

183

∂f |
cte. uγ v2γ e−γ |a| . ∂ ua

(and similar for the derivative with respect to va ). The Hessian of f verifies |

∂2 f |
cte. uγ vγ e−γ |a−c| , ∂ ua ∂ vc

and |

∂2 f |
cte. v2γ e−γ |a+b| ∂ ua ∂ ub

(and similar for the second derivative with respect to vc , vd ). The exponential decay of the second derivatives can of course be formulated in the real coordinates (ξ , η ) but the description is again more complicated (see Sect. 5.2).

1.4 Action-angle variables Let A be a finite subset of Zd and fix 0 < pa ,

a∈A.

The (#A )-dimensional torus 1 2 2 2 (ξa + ηa ) =

pa , a ∈ A

ξa = ηa = 0,

a ∈ L = Zd \ A ,

is invariant for the Hamiltonian flow when ε = 0. In the symplectic subspace RA × RA we introduce, in a neighborhood of this torus, action-angle variables (ra , ϕa ), a∈A,  ξa = 2(pa + ra ) cos(ϕa )  ηa = 2(pa + ra ) sin(ϕa ). These coordinates are analytic near r = 0 because the pa ’s are all positive. In these coordinates the Hamiltonian equations becomes ⎧ ⎪ ζ = J ∂∂ζa (h + ε f ) a ∈ L ⎪ ⎨ a r˙a = − ∂ ∂ϕa (h + ε f ) ⎪ ⎪ ⎩ ϕ˙ = ∂ (h + ε f ) a ∈ A a

∂ ra

with the integrable part h(ξ , η , r) =

1

∑ ωa ra + 2 ∑ Ωa (ξa2 + ηa2 )

a∈A

a∈L

184

L.H. Eliasson and S.B. Kuksin

(modulo a constant), where

ωa = |a|2 + Vˆ (a),

a∈A,

are the basic frequencies, and

Ωa = |a|2 + Vˆ (a),

a∈L,

are the normal frequencies (of the invariant torus). The perturbation ε f (ξ , η , r, ϕ ) will be a function of all variables (under the assumption, of course, that the torus lies in the domain of F). Since h + ε f is analytic in (some domain in) the phase space l02 (L , R2 ) × RA × A T , it extends to a holomorphic function on a complex domain ⎧  ⎪ ⎨ ζ 0 = ξ 20 + η 20 < σ O 0 (σ , µ , ρ ) = |r| < µ ⎪ ⎩ | Im ϕ | < ρ .

1.5 Statement of the result The Hamiltonian h + ε f is a standard form for the perturbation theory of lowerdimensional (isotropic) tori with one exception: it is strongly degenerate. We therefore need external parameters to control the basic frequencies and the simplest choice is to let the basic frequencies (i.e. the potential itself) be our free parameters. The parameters will belong to a set U ⊂ {ω ∈ RA : |ω |
C1 } .

(3)

The potential V will be analytic and |Vˆ (a)|
C2 e−C3 |a| , C3 > 0, ∀a ∈ L . The normal frequencies will be assumed to verify ⎧ ⎨ |Ωa |  C4 > 0 |Ωa + Ωb |  C4 ∀ a, b ∈ L . ⎩ |Ωa − Ωb |  C4 |a| = |b|

(4)

(5)

This is fulfilled, for example, if V is small and A  0, or if V is arbitrary and A is sufficiently large. Theorem 1.1. Under the above assumptions, for ε sufficiently small there exist a subset U ⊂ U, which is large in the sense that Leb (U \U )
cte.ε exp ,

Four lectures on KAM for the non-linear Schrödinger equation

185

and for each ω ∈ U , a real analytic symplectic diffeomorphism Φ

σ µ ρ O 0 ( , , ) → O 0 (σ , µ , ρ ) 2 2 2 and a vector ω = ω (ω ) such that (hω + ε f ) ◦ Φ equals (modulo a constant) 1 + +ε g , 2 where g ∈ O(|r|2 , |r| ζ 0 , ζ 30 ) and the symmetric matrix A(ω ) has the form   Ω1 (ω ) Ω2 (ω ) t Ω (ω ) Ω (ω ) 2 1 with Ω1 + iΩ2 Hermitian and block-diagonal, with finite-dimensional blocks. Moreover, Φ = (Φζ , Φr , Φϕ ) verifies, for all (ζ , ϕ , r) ∈ O 0 ( σ2 , µ2 , ρ2 ), 0 0   0Φζ − ζ 0 + |Φr − ρ | + Φϕ − ϕ 
β ε , 0 and the mapping ω → ω (ω ) verifies   ω − id
β ε. Lip(U ) The exponent exp only depends on the dimensions d, #A , m∗ , the constant cte. depends on the dimensions and on C1 , . . . ,C4 , and the constant β also depends on V and F. It follows from this theorem that Φ ({0} × {0} × TA ) is a KAM-torus for the Hamiltonian system of h + ε f , and it implies the result mentioned in Sect. 1.1. We discuss this notion and its consequences in the next section. Theorem 1.1, as well as a more generalized version, is proven in [EK06].

1.6 KAM-tori A KAM-torus of a Hamiltonian system in R2L × RA × TA is a finite-dimensional torus satisfying (i) Invariance – it is invariant under the Hamiltonian flow (ii) Linearity – the flow on the torus is conjugate to a linear flow ϕ → ϕ + t ω A torus with the two properties (i) + (ii) is nothing more and nothing less than a quasi-periodic solution when translated into cartesian coordinates. Often, as we shall do in this paper, one also requires

186

L.H. Eliasson and S.B. Kuksin

(iii) Reducibility – the linearized equations (the “variational equations”) on the torus are conjugate to a constant coefficient system of the form ⎧ d ζˆ ⎪ ⎪ = JAζˆ ⎪ ⎨ dt d rˆ dt = 0 ⎪ ⎪ ⎪ ⎩ d ϕˆ dt = β rˆ and JA has a pure point spectrum If the quasi-periodic solution has property (iii), then questions related to linear stability and Lyapunov exponents “reduce” to a study of a linear system of constant coefficients, which permits (at least for finite-dimensional systems) to answer such questions and to construct higher order normal forms near the torus. Reducibility is automatic in two cases: if the torus is one-dimensional (and phasespace is finite-dimensional) it is just a periodic solution, and (iii) is a general fact called Floquet theory; if the torus is Lagrangian (i.e. there is not ζ -part), then (iii) follows from (i) + (ii) by a simple integration [dlL01]. In general, however, it is a delicate property which is far from being completely understood. KAM is a perturbation theory of KAM-tori. Not only is reducibility an important outcome but also an essential ingredient in the proof. It simplifies the iteration since it reduces all approximate linear equations to constant coefficients. But it does not come for free. It requires a lower bound on small divisors of the form (∗∗)

| +Ωa (ω ) ± Ωb (ω )| ,

k ∈ ZA , a, b ∈ L ,

where Ωa (ω ), a ∈ L are the imaginary parts of the eigenvalues of JA(ω ) The basic frequencies ω will be fixed during the iteration – that is what parameters are there for – but the normal frequencies will vary. Indeed the Ωa (ω ) are perturbations of |a|2 + Vˆ (a) which are not known a priori but are determined by the approximation process.1 The difficulty associated with the small divisors (∗∗) may be very large. There is a perturbation theory which avoids this difficulty, but to a high cost: the approximate linear equations are no longer of constant coefficients. Moreover it gives persistence of the invariant tori but no reducibility.

1.7 Consequences of Theorem 1 The consequences of the theorem is that Φ ({0, 0} × TA ) is a KAM-torus for hω + ε f . In order to see this it suffices to show that {ζ = r = 0} is a KAM-torus for k + ε g, 1 A lower bound on (∗∗) is strictly speaking not necessary for reducibility. It is necessary, however, in order to have reducibility with a reducing transformation close to the identity.

Four lectures on KAM for the non-linear Schrödinger equation

187

1 k = + . 2 Since

∂g ∂g ∂g =0 = = ∂ζ ∂ϕ ∂r

for ζ = r = 0, it follows that {ζ = r = 0} is invariant with a flow ϕ → ϕ + t ω . The linearized equations on this torus become ⎧ ˆ dζ ⎪ ⎪ = JA(ω )ζˆ + ε Ja(ϕ + t ω , ω )ˆr ⎪ ⎪ ⎪ ⎨ dt d rˆ =0 ⎪ dt ⎪ ⎪ ⎪ ⎪ ⎩ d ϕˆ = ε +ε b(ϕ + t ω , ω )ˆr dt where2 a(ϕ ) = ∂ ∂r∂ ζ g(0, 0, ϕ ) and b(ϕ ) = ∂∂r2 g(0, 0, ϕ ). These equations can be conjugated to constant coefficients if the imaginary part of the the eigenvalues of JA(ω ), 2

2

±iΩa (ω ),

a∈L,

are non-resonant with respect to ω . In order to see this we consider the equations (i) = JAZ1 (ϕ ) + ε Ja(ϕ ), which has a unique smooth solution if ω and Ωa (ω ), a ∈ L , verify an appropriate Diophantine condition (ii) = −Z2 (ϕ )JA + ε ta(ϕ ) which has a unique smooth solution under the same condition on ω (iii) = ε ta(ϕ )Z1 (ϕ ) + ε b(ϕ ) − εβ which has a smooth solution if ω is Diophantine and if we chose β such that the meanvalue of the right hand side is = 0. If we now take

2

⎞ I Z1 (ϕ ) 0 I 0 ⎠, Z(ϕ ) = ⎝ 0 Z2 (ϕ ) Z3 (ϕ ) I ⎛

t is used both as the independent time-variable and to denote transposition, without confusion we hope.

188

L.H. Eliasson and S.B. Kuksin

then < ∂∂ ϕZ (ϕ ), ω > ⎞ ⎞ ⎛ JA 0 0 JA ε Ja(ϕ ) 0 0 0 ⎠ Z(ϕ ) − Z(ϕ ) ⎝ 0 0 0 ⎠ , =⎝ 0 ε ta(ϕ ) ε b(ϕ ) 0 0 εβ 0 ⎛

so Z conjugates the linearized equations to ⎧ ⎪ d ζˆ ⎪ ⎪ = JA(ω )ζˆ ⎪ ⎪ ⎨ dt d rˆ =0 ⎪ dt ⎪ ⎪ ⎪ d ϕˆ ⎪ ⎩ = εβ rˆ dt which is constant coefficients. The conditions on ω will hold if we restrict the set U arbitrarily little. If   1 I I , C= √ 2 −iI iI then C−1 JA(ω )C = i

t

(6)

 Ω (ω ) 0 , 0 −Ω (ω )

since Ω (ω ) = Ω1 (ω ) + iΩ2 (ω ) is Hermitian. Moreover, there is a unitary matrix D = D(ω ) such that t ¯ DΩ D = diag(Ωa ) is a real diagonal matrix, and therefore 

D 0 0 D¯

−1  t     D 0 diag(Ωa ) Ω 0 0 = i i 0 D¯ 0 −Ω 0 −diag(Ωa )

So the linearized equations on the torus have only quasi-periodic solutions and, hence, the torus is linearly stable.

1.8 References For finite dimensional Hamiltonian systems the first proof of persistence of stable (i.e. vanishing of all Lyapunov exponents) lower dimensional invariant tori was obtained in [Eli85, Eli88] and there are now many works on this subjects. There are also many works on reducibility (see for example [Kri99, Eli01]) and the situation in finite dimension is now pretty well understood in the perturbative setting. Not so, however, in infinite dimension.

Four lectures on KAM for the non-linear Schrödinger equation

189

If d = 1 and the space-variable x belongs to a finite segment supplemented by Dirichlet or Neumann boundary conditions, this result was obtained in [Kuk88] (also see [Kuk93, Pös96]). The case of periodic boundary conditions was treated in [Bou96], using another multi-scale scheme, suggested by Fröhlich–Spencer in their work on the Anderson localization [FS83]. This approach, often referred to as the Craig–Wayne scheme, is different from the KAM-scheme described here. It avoids the cumbersome condition (∗∗) but to a high cost: the approximate linear equations are not of constant coefficients. Moreover, it gives persistence of the invariant tori but no reducibility and no information on the linear stability. A KAMtheorem for periodic boundary conditions has recently been proved in [GY05] (with a perturbation F independent of x) and the perturbation theory for quasi-periodic solutions of one-dimensional Hamiltonian PDE is now sufficiently well developed (see for example [Kuk93, Cra00, Kuk00]). The study of the corresponding problems for d  2 is at its early stage. Developing further the scheme, suggested by Fröhlich–Spencer, Bourgain proved persistence for the case d = 2 [Bou98]. More recently, the new techniques developped by him and collaborators in their work on the linear problem has allowed him to prove persistence in any dimension d [Bou04]. (In this work he also treats the non-linear wave equation.) For another, and simplified, proof of this result see [Yua07].

1.9 Notation is the standard scalar product in Rd .   is an operator-norm or l 2 -norm. | | will in general denote a supremum norm, with a notable exception: for a lattice vector a ∈ Zd we use |a| for the l 2 -norm. A is a finite subset of Zd , and L is its complement. A matrix on L is just a mapping A : L × L → C or gl(2, C). Its components will be denoted Aba . If A1 , A2 , A3 , A4 are scalar-valued matrices on L , then we identify   A1 A2 A= A3 A4 with a gl(2, C)-valued matrix through   (A1 )ba (A2 )ba b . Aa = (A3 )ba (A4 )ba The dimension d will be fixed and m∗ will be a fixed constant > d2 .  means
modulo a multiplicative constant that only, unless otherwise specified, depends on d, m∗ and #A . The points in the lattice Zd will be denoted a, b, c, . . . . Also d will sometimes be used, without confusion we hope.

190

L.H. Eliasson and S.B. Kuksin

For two subsets X and Y of a metric space, dist(X,Y ) =

inf d(x, y).

x∈X,y∈Y

(This is not a metric.) Xε is the ε -neighborhood of X, i.e. {y : dist(y, X) < ε }. Let Bε (x) be the ball {y : d(x, y) < ε }. Then Xε is the union, over x ∈ X, of all Bε (x). If X and Y are subsets of Rd or Zd we let X −Y = {x − y : x ∈ X, y ∈ Y } – not to be confused with the set theoretical difference X \Y .

2 The homological equation Here we shall describe shortly the quadratic iteration and derive the homological equation which is central in KAM.

2.1 Normal form Hamiltonians This is a real Hamiltonian of the form 1 h = + , 2 

where A=

Ω1 Ω2 tΩ Ω 2 1

(modulo a constant) 

is block-diagonal matrix with finite-dimensional blocks (we shall say more about these blocks in Sect. 3) and Ω (ω ) = Ω1 (ω ) + iΩ2 (ω ) is Hermitian. Since Ω (ω ) is Hermitian the eigenvalues of JA(ω ) are ±iΩa (ω ) a ∈ L , where the Ωa (ω ) are the (necessarly real) eigenvalues of Ω (ω ). (See the discussion in Sect. 2.2.) We also suppose A(ω ) to be close to   diag(|a|2 + Vˆ (a) 0 0 diag(|a|2 + Vˆ (a))

Four lectures on KAM for the non-linear Schrödinger equation

and

191

1 ∂ω Ω (ω )
. 4

This implies that Ωa (ω ) is

≈ |a|2 + Vˆ (a)

and C 1 (or Lipschitz) -small in ω .

2.2 The KAM-iteration Given a normal form Hamiltonian 1 h = + 2 and a perturbation f . Let T f be the Taylor polynomial f (0, 0, ϕ )+
+ < (0, 0, ϕ ), ζ > + ∂r ∂ζ 2 ∂ζ

of f – it may also depend on ω . If T f was = 0 then {ζ = r = 0} would be a KAM-torus for h + f . But in general we only have T f ∈ O(ε ). Suppose now there exist a Taylor polynomial s of the same form, i.e. s = T s, and a normal form Hamiltonian 1 k = c(ω )+ + 2 verifying {h, s} = −T f + k, (1) where { , } is the Poisson bracket associated to the symplectic form ∑ d ξa ∧ d ηa + ∑ dra ∧ d ϕa . This equation is known as the homological equation. Let Φ t be the flow of ⎧˙ ζ = J ∂∂ζs (ζ , ϕ , r) ⎪ ⎪ ⎨ r˙ = − ∂∂ϕs (ζ , ϕ , r) ⎪ ⎪ ⎩ ϕ˙ = ∂∂ rs (ζ , ϕ , r). If s, k ∈ O(ε ), then (Φ t − id) ∈ O(ε ) and (h + f ) ◦ Φ 1 = h + k + = h+k+ = h+k+

1 d t 0 dt (h + t f + (1 − t)k) ◦ Φ dt 1 0

({h + t f + (1 − t)k, s} + f − k) ◦ Φ t dt

0

({t f + (1 − t)k, s} + f − T f ) ◦ Φ t dt

1

= h + k + [( f − T f ) + f1 ].

192

L.H. Eliasson and S.B. Kuksin

So Φ 1 transforms h + f to a new normal form h = h + k plus a new perturbation f . Since T ( f ) ∈ O(ε 2 ), also

f ∈ O(ε 2 )

when the domain is sufficiently restricted. If we can solve the homological equation (1), not only for the normal form Hamiltonian h but also for all normal form Hamiltonians h , close to h, then we will be able to make an iteration which will converge to a solution as in Theorem 1.1 if the estimates a good enough. So the basic thing in KAM is to solve and estimate the solution of the homological equation. It is clear from the discussion above that it is enough to solve a slightly weaker version of the homological equation, namely {h, s} = −T f + k + O(ε 2 ).

(2)

2.3 The components of the homological equation We write s as 1 S01 (ϕ ) + + + 2 and k as

1 c+ + . 2 The homological equation (2) now decomposes into four linear equations. The first two are & = − f (0, 0, ϕ ) + c + O(ε 2 ) (3) = − ∂∂ rf (0, 0, ϕ ) + χ + O(ε 2 ). In these equations, we are forced to take c = < f (0, 0, ·)> where is the mean value 1 (2π )d

and  Td

χ =
, ∂r

g(ϕ )d ϕ .

The other two are +JAS1 (ϕ ) = −

∂f (0, 0, ϕ ) + O(ε 2 ) ∂ζ

(4)

Four lectures on KAM for the non-linear Schrödinger equation

193

and +AJS2 (ϕ ) − S2 (ϕ )JA = − ∂∂ ζ 2f (0, 0, ϕ ) + B + O(ε 2 ). 2

(5)

The most delicate of these equations is the last one which is related to reducibility. Let   Ω Ω B = t 1 2 , Ω = Ω1 + iΩ2 , Ω2 Ω1 and F(ϕ ) =

∂2 f (0, 0, ϕ ). ∂ζ2

˜ ϕ ) = t CF(ϕ )C and S˜2 (ϕ ) = t CS2 (ϕ )C, then (5) becomes If we write F(     0 Ω 0 Ω J S˜2 (ϕ ) + iS˜2 (ϕ )J t −i t Ω 0 Ω 0  

˜ ϕ ) + t 0 Ω + O(ε 2 ). = −F( Ω 0 This equation decouples into four equations for scalar-valued matrices. These are of the form (6) +i(Ω R(ϕ ) + R(ϕ )t Ω ) = G(ϕ ) + O(ε 2 ), for the diagonal terms, and of the form +i(Ω R(ϕ ) − R(ϕ )Ω ) = G(ϕ ) − Ω + O(ε 2 )

(7)

for the off-diagonal terms. The last equation is underdetermined and there are several possible choices of Ω . One such choice would be which would give an Hermitian matrix, but in general not a block diagonal matrix. So the Hamiltonian h = h + k would not be on normal form. Instead we shall make a “smaller” choice. Due to the exponential decay of the second order derivatives of the Hamiltonian (discussed in Sect. 1.3) the matrix G verifies |G(ϕ )ba |  ε e−γ |a−b|

a, b ∈ L ,

and we can truncate the matrices away from the diagonal at distance 1 ∆ ≈ log( ). ε We then take $ if |a| = |b|, |a − b|
∆ (Ω )ba = 0 if not

(8)

Since the left hand side of the equations (3–7) are linear operators with constant coefficients, equations (3–7) + (8) can be solved in Fourier series, and to get a solution we must prove the convergence of these Fourier series and estimate the

194

L.H. Eliasson and S.B. Kuksin

solution. This requires good estimates on the small divisors, i.e. the eigenvalues of the linear operators in the left hand side.

2.4 Small divisors and the second Melnikov condition Since the equations are to be solved only modulo O(ε 2 ) and since all functions are analytic in ϕ , we can truncate all Fourier series at order   1

. ∆ ≈ log ε We want to bound the eigenvalues (in absolute value) in the left hand side from below by some quantity κ which should be small but much larger than ε , say

κ = ε exp for some small exponent. For (3), the eigenvalues of the left hand side operator are i

k ∈ ZA , 0 < |k|
∆ .

These are all larger (in absolute value) than κ for ω ∈ U except on a small set of Lebesgue measure  (∆ )#A κ . The eigenvalues in (4) are i < k, ω > +iΩa (ω ) k ∈ ZA , |k|
∆ ,

a∈L,

where the iΩa (ω ):s are the eigenvalues of A(ω ). By the assumption on A(ω ),

Ωa (ω ) ≈ |a|2 + Vˆ (a) and is C 1 -small in ω . Therefore there are only finitely many eigenvalues which are not large, and these can be controlled by an appropriate choice of ω . Equation (6) is treated in the same way. It is (7) which gives rise to serious problems. If we define Ω by (8) and take into account the exponential decay of the matrices, then the eigenvalues of (7) are i(Ωa (ω ) − Ωb (ω )) k = 0, |a − b|
∆ , |a| = |b|, (which are all  1 by assumption (5) of Section 1) and $ i +Ωa (ω ) − Ωb (ω )) 0 < |k|
∆ , |a − b|
∆ . In one space dimension d = 1 we have

(9)

Four lectures on KAM for the non-linear Schrödinger equation

195

|Ωa (ω ) − Ωb (ω )| → ∞ when |a| → ∞, |a − b|
∆ , except for a = b. Therefore there are only finitely many eigenvalues which are not large, and these can be controlled by an appropriate choice of ω . But in dimension d  2 there are infinitey many eigenvalues which are not large. How to control (9) – known as the second Melnikov condition – is the main difficulty in the proof. But before we turn to this question we shall discuss more closely the normal form.

3 Normal form Hamiltonians We shall discuss the block-diagonal property and the Töplitz–Lipschitz-property of the normal form Hamiltonians.

3.1 Blocks In this section d  2. For a non-negative integer ∆ we define an equivalence relation on L generated by the pre-equivalence relation $ 2 |a| = |b|2 a ∼ b ⇐⇒ |a − b|
∆. Let [a]∆ denote the equivalence class (block) of a, and let E∆ be the set of equivalence classes. It is trivial that each block [a] is finite with cardinality  |a|d−1 that depends on a. But there is also a uniform ∆-dependent bound. Lemma 1 Let d∆ = sup(diam[a]∆ ). a

Then d∆  ∆

(d+1)! 2

.

Proof. We give the proof in dimension d = 2, the general case being treated in Sect. 4 of [EK06]. It suffices to consider the case when there are a, b, c ∈ [a]∆ such that a − b and a − c are linearly independent and |a − b|, |a − c|
∆.

196

L.H. Eliasson and S.B. Kuksin

(If not, then [a]∆ = {a, b} and the result is obvious.) Since |a|2 = |b|2 = |c|2 it follows that $ = 12 |a − b|2 = 12 |a − c|2 Since a − b and a − c are integer-valued independent vectors it follow from this that |a|  ∆3 .



The blocks [a]∆ have a rigid structure when |a| is large. For a vector c ∈ Zd \ 0 let ac ∈ (a + Rc) ∩ Zd be the lattice point b on the line a + Rc with smallest norm – if there are two such b’s we choose the one with  0. Lemma 2 Given a and c = 0 in Z∆ . For all t, such that |a + tc|  d∆2 (|ac | + |c|) |c| , the set [a + tc]∆ − (a + tc) is independent of t and ⊥ to c. Proof. It suffices to prove this for a = ac . Let b ∈ [a + tc]∆ − (a + tc) for some fixed t as in the lemma. This implies that |b|
d∆ and that |b + a + tc|2 = |a + tc|2 . This last equality can be written 2t +2 +|b|2 = 0. If = 0, then |a + tc|
|a| + |t ||c| = |a| + | ||c| + 12 |b|2 ||c|
(1 + d∆ )|a||c| + 12 d∆2 |c|2 ), but this is impossible under the assumption on a + tc. Therefore = 0 and hence [a +tc]∆ − (a +tc) ⊥ c. Moreover |b + a + sc|2 = |a + sc|2 for all s, so if |b|
∆, then [b + a + sc]∆ = [a + sc]∆

∀s.

To conclude, let b0 = a, b1 , . . . , bn be the elements of [a]∆ ordered in such a way that |b j+1 − b j |
∆ for all j. Then the preceding argument shows that [b + a + sc]∆ = [a + sc]∆

∀s, ∀ j.



Description of blocks when d = 2, 3. For d = 2, we have outside {|a| :
d∆ ≈ ∆3 }  Rank[a]∆ = 1 if, and only if, a ∈ {a, a − b}

b 2

+ b⊥ for some 0 < |b|
∆ – then [a]∆ =

Four lectures on KAM for the non-linear Schrödinger equation

197

 Rank[a]∆ = 0 otherwise – then [a]∆ = {a} For d = 3, we have outside {|a| :
d∆ ≈ ∆12 }  Rank[a]∆ = 2 if, and only if, a ∈ b2 + b⊥ ∩ 2c + c⊥ for some 0 < |b| , |c|
2∆ linearly independent – then [a]∆ ⊃ {a, a − b, a − c}  Rank[a]∆ = 1 if, and only if, a ∈ b2 + b⊥ for some 0 < |b|
∆ – then [a]∆ = {a, a − b}  Rank[a]∆ = 0 otherwise – then [a]∆ = {a}

3.2 Lipschitz domains For a non-negative constant Λ and for any c ∈ Zd \ 0, let the Lipschitz domain DΛ (c) ⊂ L × L be the set of all (a, b) such that there exist a , b ∈ Zd and t  0 such that &

|a = a + tc|  Λ(|a | + |c|) |c| |b = b + tc|  Λ(|b | + |c|) |c|

and

|a| , |c|

|b|  2Λ2 . |c|

The Lipschitz domains are not so easy to grasp, but it is easy to verify Lemma 3 Let Λ  3. (i) If |a = a + t0 c|  Λ(|a | + |c|)|c|, t  0, then |a| ≈ ≈ t  Λ|c|. |c| |c|2 (ii) If |a = a + t0 c|  Λ(|a | + |c|)|c|, t0  0, then |a + tc|2  |a + t0 c|2 + (t − t0 )2 |c|2

∀t  t0 .

In particular, if (a, b) ∈ DΛ (c), then (a + tc, b + tc) ∈ DΛ (c) ∀t  0. Proof. (i) The inequality |a + tc|
|a | + t|c|
(|a | + t)|c| gives immediately that t  Λ|c|. It also gives Λ(|a | + |c|)
|a | + t,

198

L.H. Eliasson and S.B. Kuksin

which implies that |a |
Since |

t . Λ−1

|a| |a | − t|, | − t|
2 |c| |c| |c|

we are done. (ii) Let s = t − t0 . Then |a + sc|2 = |a|2 + s2 |c|2 + 2s and 2s = 2st0 |c|2 + 2s  2st0 (|c|2 − which is  0.

|a ||c| ) t0



A bit more complicated is Lemma 4 For any |a|  Λ2d−1 , there exist c ∈ Zd , 0 < |c|  Λd−1 , such that |a|  Λ(|ac | + |c|) |c| ,  0. Proof. For all K  1 there is a c ∈ Zd ∩ {|x|
K} such that 1 1 d = dist(c, Ra)
C1 ( ) d−1 K where C1 only depends on d. K a] in Rd and a tubular neighborTo see this we consider the segment Γ = [0, |a| hood Γε of radius ε : vol(Γε ) ≈ K ε d−1 . The projection of Rd onto Td is locally injective and locally volume-preserving. If 1 ε  ( K1 ) d−1 , then the projection of Γε cannot be injective (for volume reasons), so there are two different points x, x ∈ Γε such that x − x = c ∈ Z d \ 0. Then |ac | 

|a| d. |c|

Now Λ(|ac | + |c|) |c|
2ΛK 2 +C2 If we choose K = (2C2 Λ)d−1 , then this is
|a|.

Λ 1

K d−1



|a| .

Four lectures on KAM for the non-linear Schrödinger equation

199

The most important property is that finitely many Lipschitz domains cover a “neighborhood of ∞” in the following sense. Corollary 3.1 For any Λ, N > 1, the subset {|a| + |b|  Λ2d−1 } ∩ {|a − b|
N} ⊂ Zd × Zd is contained in

/

Dω (c)

0Λ,γ = LipΛ,γ X + |X|γ . We say that the matrix X is Töplitz–Lipschitz if < X >Λ,γ < ∞ for some Λ, γ .

|a| |b| γ |a−b| , )e |c| |c|

Four lectures on KAM for the non-linear Schrödinger equation

ˆ Example 2 Consider R(k) from the example above. If (a, b) ∈ DΛ (c),

Λ  3,

then |a = a + tc|
Λ(|a | + ||c|)|c| and By Lemma 3 we have

|b = b + tc|
Λ(|b | + ||c|)|c|.

|a| |b| ≈ ≈ t  Λ. |c| |c|

If = 0 then   |a| |b| ˆ b  R(k)a − 0 max( , )eγ |a−b| |c| |c|     ˆ ba G(k)  γ |a−b|  ≈ e 1  + t ( +|a |2 − |b |2 + Vˆ (a) − Vˆ (b))  which is

  ≈ 

 ˆ ba  γ |a−b| G(k) e  |G|γ 

if Λ, hence t, is sufficiently large. If = 0 then   |a| |b|  ˆ b ˆ R(k)a − R(k)(c)ba  max( , )eγ |a−b| |c| |c|    2       1 1     G(k) ˆ  ˆ  LipΛ,γ (G(k)) +    γ

2

2

2

2 +|a | − |b | +|a | − |b | if Λ, hence t, is sufficiently large. Here we have used the decay of Vˆ to bound |Vˆ (a + tc) − Vˆ (b + tc)|t  1. ˆ ˆ In particular, the matrix R(k) is Töplitz–Lipschitz if G(k) is Töplitz–Lipschitz.

3.5 Normal form Hamiltonians Consider the class of Hamiltonians 1 h = + , 2

(modulo a constant)

201

202

L.H. Eliasson and S.B. Kuksin

from Sect. 2.1. We say that h is N F ∆ if moreover Ω is block-diagonal over E∆ , i.e. Ωab = 0 if [a]∆ = [b]∆ Clearly if h is N F ∆ for some ∆
∆ , then, by the choice of Ω , in (8) of Section 2 h = h + k is N F ∆ , where 1 k = c+ + 2 is determined in Sect. 3.2. Let H(ω ) = Ω (ω ) − diag(|a|2 + Vˆ (a) : a ∈ L ). We shall also require that H(ω ) and ∂ω H(ω ) are Töplitz at i for all ω ∈ U and uniformly Töplitz–Lipschitz, i.e. there is a Λ such that < H >& Λ 1 = sup(< H(ω >Λ , < ∂ω H(ω >Λ ) < ∞). U

U

Then, clearly, the convergence to the Töplitz-limits is uniform in ω both for (H(ω ) and ∂ω H(ω ).

4 Estimates of small divisors Here we verify the second Melnikov condition for the normal form Hamiltonians described in Sect. 3.5.

4.1 A basic estimate Lemma 5 Let f : I =] − 1, 1[→ R be of class C n and    (n)   f (t)  1 ∀t ∈ I. Then, ∀ε > 0, the Lebesgue measure of {t ∈ I : | f (t)| < ε } is 1


cte.ε n , where the constant only depends on n.   0   Proof. We have  f (n) (t)  ε n for all t ∈ I. Since f (n−1) (t) − f (n−1) (t0 ) =

 t t0

f (n) (s)ds,

Four lectures on KAM for the non-linear Schrödinger equation

203

  1 1   we get that  f (n−1) (t)  ε n for all t outside an interval of length
2ε n . By induc  j 1   tion we get that  f (n− j) (t)  ε n for all t outside 2 j−1 intervals of length
2ε n . j = n gives the result.

 Remark 1 The same is true if     max  f ( j) (t)  1 ∀t ∈ I

0
j
n

and f ∈ C n+1 . In this case the constant will depend on | f |C n+1 . Let A(t) be a real diagonal N × N-matrix with diagonal components a j which are C 1 on I =] − 1, 1[ and a j (t)  1

j = 1, . . . , N, ∀t ∈ I.

Let B(t) be a Hermitian N × N-matrix of class C 1 on I =] − 1, 1[ with 0 0 1 0B (t)0
2

∀t ∈ I.

Lemma 6 The Lebesgue measure of the set {t ∈ I :

min

λ (t)∈σ (A(t)+B(t))

|λ (t)| < ε }

is
cte.N ε , where the constant is independent of N. Proof. Assume first that A(t) + B(t) is analytic in t. Then each eigenvalue λ (t) and its (normalized) eigenvector v(t) are analytic in t, and

λ (t) =< v(t), (A (t) + B (t))v(t) (scalar product in CN ). Under the assumptions on A and B, this is  1 − 12 . Lemma 5 applied to each eigenvalue λ (t) gives the result. If B is non-analytic we get the same result by analytic approximation.

 We now turn to the main problem.

4.2 The second Melnikov condition (d = 2) Proposition 7 Let ∆ > 1 and 0 < κ < 1. Assume that U verifies (3) of Section 1, that Ω = diag(|a|2 + Vˆ (a) : a ∈ L )

204

L.H. Eliasson and S.B. Kuksin

verifies (4) of Section 1 and that H : L × L → C verifies ∂ω H(ω )


1 4

ω ∈ U.

(1)

(  is the operator norm.) Assume also that H(ω ) and ∂ω H(ω ) are Töplitz at ∞ and N F ∆ for all ω ∈ U. Then there exists a subset U ⊂ U, Leb(U \U )
1 cte. max(∆ , d∆2 , Λ)exp+#A (C1 + < H >& Λ 1 )2 κ 3 C1#A −1 , U such that, for all ω ∈ U , 0 < |k|
∆ and all dist([a]∆ , [b]∆ )
∆

(2)

we have $ | +α (ω ) − β (ω )|  κ

∀

α (ω ) ∈ σ ((Ω + H)(ω )[a]∆ ) β (ω ) ∈ σ ((Ω + H)(ω )[b]∆ ).

(3)

Moreover the κ -neighborhood of U \U satisfies the same estimate. The exponent exp is a numerical constant. The constant cte. depends on #A and on C2 ,C3 . Proof. The proof goes in the following way: first we prove an estimate in a large finite part of L (this requires parameter restriction); then we assume an estimate “at ∞” of L and we prove, using the Lipschitz-property, that this estimate propagate from “∞” down to the finite part (this requires no parameter restriction); in a third step we have to prove the assumption at ∞. Let us notice that it is enough to prove the statement for ∆  max(Λ, d∆2 ). We let [ ] denote [ ]∆ . For each k, [a]∆ , [b]∆ it follows by Lemma 6 the set of ω such that | +α (ω ) − β (ω )| < κ has Lebesgue measure  d∆4

κ #A −1 C . |k| 1

1. Finite part. For the finite part, let us suppose a belongs to ⎞ ⎛ 1 ⎟ ⎜ {a ∈ L : |a|  ⎝C1 + d∆2 < H >& Λ 1 ⎠ (∆ )6 }, κ1 U

(4)

Four lectures on KAM for the non-linear Schrödinger equation

205

1

where3 κ1 = κ 3 . These are finitely many possibilities and (3)κ is fulfilled, for all [a] satisfying (4), all [b] with |a − b|  ∆ and all 0 < |k|
∆ , outside a set of Lebesgue measure  (C1 + d∆2 < H >& Λ 1 )2 (∆ )12 (∆ )2+#A −1 d∆4 U

κ #A −1 C . κ12 1

(5)

Let us now get rid of the diagonal terms Vˆ (a, ω ) = Ωa (ω ) − |a|2 which, by (4), are
C2 e−|a|C3 . We include them into H. Since they are diagonal, H will remain on normal form. Due to the exponential decay of Vˆ , H and ∂ω H will remain Töplitz at ∞. The Lipschitz norm gets worse but this is innocent in view of the estimates. Also the estimate of ∂ω H(ω ) gets worse, but if a is outside (4) then condition (1) remains true with a slightly worse bound, say 3 ∂ω H(ω )
, 8

ω ∈ U.

So from now on, a is outside (4) and

Ωa = |a|2 . 2. Condition at ∞. For each vector c ∈ Zd such that 0 < |c|  (∆ )2 , we suppose that the Töplitz limit H(c, ω ) verifies (3)κ1 for (2) and for ([a] − [b]) ⊥ c.

(6)

It will become clear in the next part why we only need (3)κ1 and (2) under the supplementary restriction (6). 3. Propagation of the condition at ∞. We must now prove that for |b − a|  ∆ and an a ∈ L outside (4), (3)κ is fulfilled. By the Corollary 3.1 we get (a, b) ∈

/

D2∆ (c).

0= 0 or not. Consider for t  0 a pair of continuous eigenvalues $ αt ∈ σ ((Ω + H(ω ))[a+tc] ) βt ∈ σ ((Ω + H(ω ))[b+tc] ) Case I: = 0. Here (Ω + H(ω ))[a+tc] X − X(Ω + H(ω ))[b+tc] equals (|a|2 + H(ω ))[a+tc] X − X(|b|2 + H(ω ))[b+tc] – the linear and quadratic terms in t cancel! By continuity of eigenvalues, lim (αt − βt ) = (α∞ − β∞ ),

t→∞

where

&

α∞ ∈ σ ((|a|2 + H(c, ω ))[a] ) β∞ ∈ σ ((|b|2 + H(c, ω ))[b] )

Since [a] and [b] verify (6), our assumption on H(c, ω ) implies that (α∞ − β∞ ) verifies (3)κ1 . For any two a, a ∈ [a] we have |a | |a| = . |c| |c| Hence

0 0 0H(ω )[a] − H(c, ω )[a] 0 |a|  d d < H >& 1 , ∆ Λ |c| U

because ∆ ≥ max(Λ, d∆ ), and the same for [b]. Recalling that a and, hence, b violate (4) this implies 0 0 0H(ω )[d] − H(c, ω )[d] 0
κ1 , d = a, b. 4

Four lectures on KAM for the non-linear Schrödinger equation

207

By Lipschitz-dependence of eigenvalues (of Hermitian operators) on parameters, this implies that κ1 |(α0 − β0 ) − (α∞ − β∞ )|
2 and we are done. Case II: = 0. We write a = ac + τ c, where ac is the lattice point on the line a + Rc with smallest norm – if there are two such points we choose the one with  0. Since |a|  2∆ (|ac | + |c|) |c| , it follows that |ac |


1 |a| . 2∆ |c|

Now, α0 − β0 differs from |a|2 − |b|2 by at most 2 H(ω )  d∆2 < H >& Λ 1 , U and |a|2 − |b|2 = −2τ −2 − |b − a|2 . Since | |  1 it follows that

τ
|α0 − β0 | + |ac |∆ + (∆ )2 + d∆2 < H >& Λ 1 . U If now |α0 − β0 |  C1 ∆ then |a|
|ac | + |τ c| is
|ac | (∆ + 1) |c| +C1 (∆ )2 |c| + d∆2 < H >& Λ 1 |c| U 1
|a| +C1 (∆ )2 |c| + d∆2 < H >& Λ 1 |c|. 2 U Since a violates (4) this is impossible. Therefore |α0 − β0 |  C1 ∆ and (3)κ holds. Hence, we have proved that (3)κ holds for any $ ' (a, b) ∈ 0& Λ 1
< H >& Λ 1 . U

U

Clearly G(ω ) is Hermitian and, by Lemma 2, G(ω ) and ∂ω G(ω ) are block diagonal over E∆ , i.e. G(ω ) and ∂ω G(ω ) are N F ∆ . Moreover G is Töplitz in the direction c1 , b 1 Gb+tc a+tc1 = Ga ,

∀a, b,tc1 .

We want to prove that G verifies (3)κ1 for all (a, b) ∈ (2) + (6)c1 , i.e. for all |a − b|  ∆

and

([a] − [b]) ⊥ c1 .

Since G is Töplitz in the direction c1 it is enough to show this for      .  |c1 |  But then all divisors are large except finitely many which we can treat as above.

(7) 

5 Functions with the Töplitz–Lipshitz property (d = 2) We discuss here shortly some other aspects related to the proof of Theorem 1.1.

5.1 Töplitz structure of the Hessian The quadratic differential

∂ζ2

=

∑

,

a,b∈L

where A : L × L → gl(2, R) is a gl(2, R)-valued matrix. It is uniquely determined by the symmetry condition t b Aa = Aab . Its properties are best seen in the complex variables  b b Pa Qa t b . ( CAC)a = Qab P¯ab Consider for example the Schrödinger equation with a cubic potential. Then

Four lectures on KAM for the non-linear Schrödinger equation

2 Paa12 =

∑

2

b1 , b2 ∈ A b1 + b2 = a1 + a2

and Qba22 =

∑

a1 , b1 ∈ A a1 − b1 = a2 − b2

In particular

$

209

  (pb1 + rb1 ) pb2 + rb2 e−i(ϕb1 +ϕb2 )

 8 (pa1 + ra1 )(pb1 + rb1 )ei(ϕa1 −ϕb1 ) .

P is symmetric Q is Hermitian.

Moreover Q is Töplitz, b Qb+c a+c = Qa

∀a, b, c,

and (since A is finite) its elements are zero at finite distance from the diagonal. In particular, this matrix is Töplitz–Lipschitz and has exponential decay off the diagonal a = b. P is also Töplitz–Lipschitz with exponential decay but in a different sense: b−c = Pab ∀a, b, c, Pa+c and has exponential decay off the “anti-diagonal” {a = −b}.

5.2 Töplitz–Lipschitz matrices L × L → g l ( 2 , R)) We consider the space gl(2, C) of all complex 2 × 2-matrices provided with the scalar product ¯ Tr(t AB). 

Let J=

 0 1 . −1 0

and consider the orthogonal projection π of gl(2, C) onto the subspace M = CI + CJ. For a matrix we define π A through

A : L × L → gl(2, C) (π A)ba = π Aba ,

∀a, b.

We define the supremum-norms |A|± γ =

sup (a,b)∈L ×L

|Aba |eγ |a∓b|

210

L.H. Eliasson and S.B. Kuksin

and − |A|γ = max(|π A|+ γ , |A − π A|γ ).

A is said to have a Töplitz-limit at ∞ in the direction c if, for all a, b the two limits lim Ab±tc t→+∞ a+tc

∃ = Aba (±, c).

A(±, c) are new matrices which are Töplitz/“anti-Töplitz” in the direction c, i.e. b Ab+c a+c (+, c) = Aa (+, c) and

b Ab−c a+c (−, c) = Aa (−, c).

If |A|γ < ∞, γ > 0, then

π A(−, c) = (A − π A)(+, c) = 0. We say that A is Töplitz at ∞ if all Töplitz-limits A(±, c) exist. We define the Lipschitz-constants Lip± Λ,γ A = sup

sup

c =0 (a,b)∈DΛ (c)

|(A − A(±, c))±b a | max(

|a| |b| γ |a∓b| , )e |c| |c|

and the Lipschitz-norm − + − < A >Λ,γ = max(Lip+ Λ,γ π A + |π A|γ , LipΛ,γ (I − π )A + |(I − π )A|γ ).

We say that A Töplitz–Lipschitz if < A >Λ,γ < ∞ for some Λ, γ . In Sect. 2 of [EK06] we prove the following multiplicative property. Proposition 8 Let A1 , . . . , An : L × L → C be Töplitz–Lipschitz matrices with exponential decay off-diagonal, i.e.   A j  < ∞ j = 1, . . . , n, γ > 0. γ Then A1 · · · An is Töplitz–Lipschitz and < A1 · · · An >Λ+6,γ
  (cte.)n Λ2 ( γ −1γ )(n−1)d+1 [∑1
k
n ∏ 1
j
n A j γ < Ak >Λ,γk ], j

j = k

where all γ1 , . . . , γn are = γ except one which is = γ . Notice that this estimate is not an iteration of the estimate for n = 2. Linear differential equation. Consider the linear system $d dt X = A(t)X X(0) = I.

Four lectures on KAM for the non-linear Schrödinger equation

211

where A(t) is Töplitz–Lipschitz with exponential decay. The solution verifies ∞

X(t0 ) = I + ∑

 t0  t1

n=1 0

0

...

 tn−1 0

A(t1 )A(t2 ) . . . A(tn )dtn . . . dt2 dt1 .

Using Proposition 8 we get for γ < γ < X(t) − I >Λ+6,γ  Λ2 ( γ −1γ )|t| exp(cte.( γ −1γ )d |t|α (t)) sup|s|
|t| < A(s) >Λ,γ , where

α (t) = sup |A(s)|γ . 0
|s|
|t|

Remark 2 A more general version of Töplitz–Lipschitz matrices is treated in [EK07]

5.3 Functions with the Töplitz–Lipschitz property Let O γ (σ ) be the set of vectors in the complex space lγ2 (L , C) of norm less than σ , i.e. O γ (σ ) = {ζ ∈ CL × CL : ζ γ < σ }. Our functions f : O 0 (σ ) → C will be defined and real analytic on the domain

O 0 (σ ).4

We say that f is Töplitz at ∞ if the vector ∂ζ f (ζ ) lies in l02 (L , C2 ) and the matrix

∂2 f (ζ ) ∂ζ2

is Töplitz at ∞ for all ζ ∈ O 0 (σ ). We define the norm [ f ]Λ,γ ,σ

to be the smallest C such that ⎧ | f (ζ )|
0C ∀ζ ∈ O 0 (σ ) ⎪ ⎨0

0∂ζ f (ζ )0
1 C ∀ζ ∈ O γ (σ ), ∀γ
γ , σ γ ⎪ ⎩ < ∂ 2 f (ζ ) >
1 C ∀ζ ∈ O γ (σ ), ∀γ
γ . Λ,γ ζ σ2

5.4 A short remark on the proof of Theorem 1.1 Our Hamiltonians are functions of ζ = (ξ , η ), r, ϕ and ω . We measure these functions in a norm given by • The [ ]Λ,γ ,σ -norm for ζ The space lγ2 (L , C) is the complexification of the space lγ2 (L , R) of real sequences. “real analytic” means that it is a holomorphic function which is real on O 0 (σ ) ∩ lγ2 (L , R).

4

212

L.H. Eliasson and S.B. Kuksin

• The sup-norm over a complex domain |r| < µ and | Im ϕ | < ρ • The C1 -norm in ω . In this norm we estimate the solution s, k of the homological equation (2) (described in Sect. 3.2) and the transformed Hamiltonian h + f = (h + f ) ◦ Φ 1 , where Φ 1 is the time-one-map of the Hamiltonian vector field of s. In order to carry this out we study the behavior of this norm under truncations, Poisson brackets, flows and compositions.

References [Bou96] J. Bourgain, Construction of approximative and almost-periodic solutions of perturbed linear Schrödinger and wave equations, Geo Function Analysis 6 (1996), 201–230. , Quasi-periodic solutions of Hamiltonian perturbations of 2D linear [Bou98] Schrödinger equation, Ann. Math. 148 (1998), 363–439. , Green’s function estimates for lattice Schrödinger operators and applications, [Bou04] Annals of Mathematical Studies, Princeton University Press, Princeton, NJ, 2004. [Cra00] W. Craig, Problèmes de petits diviseurs dans les équations aux dérivées partielles, Panoramas et Synthéses, no. 9, Société Mathématique de France (2000). [dlL01] R. de la Llave, An introduction to kam, Proc. Sympos. Pure Math. 69 (2001). [EK06] L. H. Eliasson and S. B. Kuksin, KAM for non-linear Schrödinger equation, Preprint (2006). [EK07] H. L. Eliasson and S. B. Kuksin, Infinite Töplitz–Lipschitz matrices and operators, ZAMP, to appear (2007). [Eli85] L. H. Eliasson, Perturbations of stable invariant tori, Report No 3, Inst. Mittag–Leffler (1985). , Perturbations of stable invariant tori, Ann. Scuola Norm. Sup. Pisa, Cl. Sci., [Eli88] IV Ser. 15 (1988), 115–147. , Almost reducibility of linear quasi-periodic systems, Proc. Sympos. Pure Math. [Eli01] 69 (2001), 679–705. [FS83] J. Fröhlich and T. Spencer, Absence of diffusion in Anderson tight binding model for large disorder or low energy, Commun. Math. Phys. 88 (1983), 151–184. [GY05] J. Geng and J. You, A KAM theorem for one dimensional Schrödinger equation with periodic boundary conditions, J. Differential Equations 209 (2005), no. 259, 1–56. [Kri99] R. Krikorian, Réductibilité des systèmes produits-croisés à valeurs dans des groupes compacts, Astérisque (1999), no. 259. [Kuk88] S. B. Kuksin, Perturbations of quasiperiodic solutions of infinite-dimensional Hamiltonian systems, Izv. Akad. Nauk SSSR Ser. Mat. 52 (1988), 41–63, Engl. Transl. in Math. USSR Izv. 32 (1989) no.1. , Nearly integrable infinite-dimensional Hamiltonian systems, Springer, Berlin [Kuk93] (1993). , Analysis of Hamiltonian PDEs, Oxford University Press, Oxford, 2000. [Kuk00] [Pös96] J. Pöschel, A KAM-theorem for some nonlinear PDEs, Ann. Scuola Norm. Sup. Pisa, Cl. Sci., IV Ser. 15 (1996) no. 23, 119–148. [Yua07] X. Yuan, A KAM theorem with applications to partial differential equations of higher dimensions, Comm. Math. Physics 275 (2007), 97–137.

A Birkhoff normal form theorem for some semilinear PDEs D. Bambusi1

Abstract In these lectures we present an extension of Birkhoff normal form theorem to some Hamiltonian PDEs. The theorem applies to semilinear equations with nonlinearity of a suitable class. We present an application to the nonlinear wave equation on a segment or on a sphere. We also give a complete proof of all the results.

1 Introduction These lectures concern some qualitative features of the dynamics of semilinear Hamiltonian PDEs. More precisely we will present a normal form theorem for such equations and deduce some dynamical consequences. In particular we will deduce almost global existence of smooth solutions (in the sense of Klainerman [Kla83]) and a result bounding the exchange of energy among degrees of freedom with different frequency. In the case of nonresonant systems we will show that any solution is close to an infinite dimensional torus for times longer than any inverse power of the size of the initial datum. The theory presented here was developed in [Bam03, BG06, DS06, BDGS07, Gré06]. In order to illustrate the theory we will use as a model problem the nonlinear wave equation (1) utt − ∆ u + µ 2 u = f (u) , µ ∈ R , on a d dimensional sphere or on [0, π ] with Neumann boundary conditions. In (1), f is a smooth function having a zero of order 2 at the origin and ∆ is the Laplace Beltrami operator. The theory of Birkhoff normal form is a particular case of the theory of close to integrable Hamiltonian systems. Concerning the extension to PDEs of Hamiltonian perturbation theory, the most celebrated results are the KAM type theorems due 1

Dipartimento di Matematica dell’Università, Via Saldini 50, 20133 Milano, Italy e-mail: [email protected]

W. Craig (ed.), Hamiltonian Dynamical Systems and Applications, 213–247. c 2008 Springer Science + Business Media B.V.


213

214

D. Bambusi

to Kuksin [Kuk87], Wayne [Way90], Craig–Wayne [CW93], Bourgain [Bou98, Bou05], Kuksin–Pöschel [KP96], Eliasson–Kuksin [EK06], Yuan [Yua06]. All these results ensure the existence of families of quasiperiodic solutions, so they only describe solutions lying on finite dimensional manifolds in an infinite dimensional phase space. On the contrary the result on which we concentrate here allows one to describe all small amplitude solutions of the considered systems. The price we pay is that the description turns out to be valid only over long but finite times. A related research stream is the one carried on by Bourgain [Bou96a, Bou96b, Bou97,Bou00] who studied intensively the behavior of high Sobolev norms in close to integrable Hamiltonian PDEs. In particular he gave some lower estimates showing that in some cases high Sobolev norm can grow in an unbound way, and also some upper estimate showing that the nonlinearity can stabilize resonant systems, somehow in the spirit of Nekhoroshev’s theorem. The paper is organized as follows. First we present the classical Birkhoff normal form theorem for finite dimensional systems and we recall its proof (see Sect. 2). Then we pass to PDEs. Precisely, in Sect. 3 we first show that the nonlinear wave equation is an infinite dimensional Hamiltonian system (Sect. 3.1) and then we present the problem met in trying to extend the normal form theorem to PDEs. Subsequently we give a heuristic discussion on how to solve such difficulties (see Sect. 3.2). Then we give a precise formulation of our Birkhoff normal form theorem (Sect. 4). This part contains only the statements of the results and is split into three subsection, in the first (Sect. 4.1) we introduce the class of functions to which the theory applies and we study its properties. In the second subsection (Sect. 4.2) we give the statement of the normal form theorem and deduce the main dynamical consequences. In the third Sect. 4.3 we give the application to the considered model. Then, in Sect. 5 we give a short discussion presenting the main open problems of the domain. Finally Sect. 6 contains the proofs of all the results. The subsections of this section are independent each other. We made an effort to give a paper which is essentially self contained. We also mention that the method introduced here in order to prove the property of localization of coefficients (the property defining our class of functions) is original.

2 Birkhoff’s theorem in finite dimensions 2.1 Statement On the phase space R2n consider a smooth Hamiltonian system H having an equilibrium point at zero. Definition 2.1. The equilibrium point is said to be elliptic if there exists a canonical system of coordinates (p, q) (possibly defined only in a neighborhood of the origin) in which the Hamiltonian takes the form

A Birkhoff normal form theorem for some semilinear PDEs

215

H(p, q) := H0 (p, q) + HP (p, q) , where

n

H0 (p, q) =

∑ ωl

l=1

p2l + q2l , 2

ωl ∈ R

(1)

(2)

and HP is a smooth function having a zero of order 3 at the origin. Remark 2.1. The equations of motion of (1) have the form

∂ HP ∂ ql ∂ HP q˙l = ωl pl + ∂ pl

p˙l = −ωl ql −

(3) (4)

Since HP has a zero of order three, its gradient starts with quadratic terms. Thus, in the linear approximation the equations (3), (4) take the form p˙l = −ωl ql =⇒ q¨l + ωl2 ql = 0 q˙l = ωl pl

(5)

namely the system consists of n independent harmonic oscillators. Definition 2.2. The vector field

  ∂H ∂H , XH (p, q) := − ∂q ∂ p

(6)

is called the Hamiltonian vector field of H. Theorem 2.1. (Birkhoff) For any positive integer r ≥ 0, there exist a neighborhood U (r) of the origin and a canonical transformation Tr : R2n ⊃ U (r) → R2n which puts the system (1) in Birkhoff Normal Form up to order r, namely H (r) := H ◦ Tr = H0 + Z (r) + R (r)

(7)

where Z (r) is a polynomial of degree r + 2 which Poisson commutes with H0 , namely H0 ; Z (r) ≡ 0 and R (r) is small, i.e. |R (r) (z)| ≤ Cr zr+3 ,

∀z ∈ U (r) ;

(8)

z − Tr (z) ≤ Cr z2 ,

∀z ∈ U (r) .

(9)

moreover, one has

An inequality identical to (9) is fulfilled by the inverse transformation Tr−1 . If the frequencies are nonresonant at order r + 2, namely if

ω · k = 0 ,

∀k ∈ Zn ,

0 < |k| ≤ r + 2

(10)

216

D. Bambusi

the function Z (r) depends on the actions I j :=

p2j + q2j 2

only. Remark 2.2. The remainder R (r) is very small in a small neighborhood of the origin. In particular, it is of order ε r+3 in a ball of radius ε . It will be shown in Sect. 4.2 that in typical cases R (r) might have a relevant effect only after a time of order ε −r .

2.2 Proof The idea of the proof is to construct a canonical transformation putting the system in a form which is as simple as possible. More precisely one constructs a canonical transformation pushing the non normalized part of the Hamiltonian to order four followed by a transformation pushing it to order five and so on. Each of the transformations is constructed as the time one flow of a suitable auxiliary Hamiltonian function (Lie transform method). We are now going to describe more precisely this method. Definition 2.3. We will denote by H j the set of the real valued homogeneous polynomials of degree j + 2. Remark 2.3. Let g ∈ H j be a homogeneous polynomial, then there exists a constant C such that (11) |g(z)| ≤ C z j+2 . The Hamiltonian vector field Xg of g is a homogeneous polynomial of degree j + 1 and therefore one has 0 0 0Xg (z)0 ≤ C z j+1 (12) with a suitable constant C . The best such that (12) holds is usually called 0 0 constant the norm of Xg and is denoted by 0Xg 0. Similarly one can define the norm of the polynomial g. Remark 2.4. If the phase space is infinite dimensional then (11) and (12) are not automatic. They hold if and only if the considered polynomial are smooth. Remark 2.5. Let f ∈ Hi and g ∈ H j then, by the very definition of Poisson Brackets one has { f , g} ∈ Hi+ j .

2.2.1 Lie transform Let χ ∈ H j be a polynomial function, consider the corresponding Hamilton equations, namely

A Birkhoff normal form theorem for some semilinear PDEs

217

z˙ = Xχ (z) , and denote by φ t the corresponding flow. Definition 2.4. The time one map φ := φ t |t=1 is called the Lie transform generated by χ . It is well known that φ is a canonical transformation. We are now going to study the way a polynomial transforms when the coordinates are subjected to a Lie transformation. Lemma 2.1. Let g ∈ Hi be a polynomial and let φ be the Lie transform generated by a polynomial χ ∈ H j with j ≥ 1. Define g0 := g ,

gl =

1 {χ ; gl−1 } , l

l≥1,

(13)

then the Taylor expansion of g ◦ φ is given by g(φ (z)) =

∑ gl (z) ,

(14)

l≥0

for all z small enough. Proof. Compute the Taylor expansion of g ◦ φ t with respect to time. Iterating the relation d g ◦ φ t = {χ , g} ◦ φ t (15) dt one has dl g ◦ φ t = {χ , ...{χ , g} ◦ φ t (16) 3 45 6 dt l l times

which gives g ◦ φt =

∑ t l gl .

(17)

l≥0

Evaluating at t = 1 one gets (14). Since Remark 2.5 implies gl ∈ Hi+l j , (14) is  the Taylor expansion of g ◦ φ as a function of the phase space variables z.

Remark 2.6. Corollary 6.1 below shows that the series (14) is convergent in a neighborhood of the origin small enough.

2.2.2 The homological equation We are now ready to construct a canonical transformation normalizing the system up to terms of fourth order. Thus let χ1 ∈ H1 be the generating function of the Lie transform φ1 , and consider H ◦ φ1 , with H given by (1). Using (14) and (13) to compute the first terms of the Taylor expansion of H ◦ φ one gets H ◦ φ = H0 + {χ1 , H0 } + H1 + h.o.t

218

D. Bambusi

where H1 is the Taylor polynomial of degree three of HP and h.o.t. denotes higher order terms. We want to construct χ1 in such a way that Z1 := {χ1 , H0 } + H1

(18)

turns out to be as simple as possible. Obviously the simplest possible form would be Z1 = 0. Thus we begin by studying the equation {χ1 , H0 } + H1 = 0

(19)

for the unknown polynomial χ1 . To study this equation define the homological operator £0 : H1 → H1 χ → £0 χ := {H0 , χ }

(20) (21)

and rewrite (19) as £0 χ1 = H1 , which is a linear equation in the finite dimensional linear space of polynomials of degree 3. Thus, if one is able to diagonalize the operator £0 ; it is immediate to understand whether the equation (19) is solvable or not. Remark 2.7. The operator £0 can be defined also on any one of the spaces H j , j ≥ 1, it turns out that £0 maps polynomials of a given degree into polynomials of the same degree. This is important for the iteration of the construction. For this reason we will study £0 in H j with an arbitrary j. It turns out that it is quite easy to diagonalize the homological operator in anyone of the spaces H j . To this end consider the complex variables 1 1 ξl := √ (pl + iql ) ; ηl := √ (pl − iql ) l ≥ 1 . 2 2

(22)

in which the symplectic form takes the form ∑l i d ξl ∧ d ηl ,1 Remark 2.8. In these complex variables the actions are given by Il = ξl ηl . 1 This means that the transformation is not canonical, however, in these variables all the theory remains unchanged except for the fact that the equations of motions have to be substituted by

∂H ξ˙l = i , ∂ ηl

η˙ l = −i

∂H , ∂ ξl

and therefore the Poisson Brackets take the form   ∂g ∂ f ∂g ∂ f { f , g} := i ∑ . − ∂ ξl ∂ ηl ∂ ηl ∂ ξl l

A Birkhoff normal form theorem for some semilinear PDEs

and H0 (ξ , η ) =

219

n

∑ ωl ξl ηl

l=1

Remark 2.9. Consider a homogeneous polynomial f of the variables (p, q), then it is a homogeneous polynomial of the same degree also when expressed in terms of the variables (ξ , η ). Remark 2.10. The monomials ξ J η L defined by

ξ J η L := ξ1J1 ξ2J2 . . . ξnJn η1L1 . . . ηnLn form a basis of the space of the polynomials. Lemma 2.2. Each element of the basis ξ J η L is an eigenvector of the operator £0 , the corresponding eigenvalue is i(ω · (L − J)). Proof. Just remark that in terms of the variables ξ , η , the action of £0 is given by

∂ f ∂ H0 ∂ f ∂ H0 −i ∂ ξ ∂ η ∂ ηl ∂ ξl l l l    ∂ ∂ f. = i ∑ ωl ηl − ξl ∂ ηl ∂ ξl l

£0 f = {H0 , f } := ∑ i

Then

ηl

∂ J L ξ η = Ll ξ J η L ∂ ηl

and thus £0 ξ J η L = iω · (L − J)ξ L η J which is the thesis.

 Thus we have that for each j the space H j decomposes into the direct sum of the kernel K of £0 and its range R. In particular the Kernel is generated by the resonant monomials, namely K = Span(ξ J η L ∈ Hi : (J, L) ∈ RS)

(23)

RS := {(J, L) : ω · (L − J) = 0}

(24)

and is the set of the resonant indexes. Obviously the range is generated by the space monomials ξ J η L with J, L varying in the complement of the resonant set. Thus it is easy to obtain the following important lemma. Lemma 2.3. Let f ∈ H j be a polynomial, write f (ξ , η ) = ∑ fJL ξ J η L J,L

(25)

220

D. Bambusi

and define Z(ξ , η ) :=

∑

fJL ξ J η L ,

χ (ξ , η ) :=

(J,L)∈RS

fJL ξ JηL i ω · (L − J) (J,L) ∈RS

∑

(26)

then one has Z = {χ , H0 } + f .

(27)

{Z, H0 } ≡ 0 .

(28)

and Motivated by the above lemma we give the following definition. Definition 2.5. A function Z will be said to be in normal form if, when written in terms of the variables ξ , η , it contains only resonant monomials, i.e. if writing Z(ξ , η ) :=

∑ ZJL ξ J η L ,

(29)

(J,L)

one has ZJL = 0 =⇒ ω · (L − J) = 0 .

(30)

Remark 2.11. A property which is equivalent to (30) is {Z, H0 } = 0, which has the advantage of being coordinate independent. Remark 2.12. If the frequencies are nonresonant, namely if eq. (10) holds, then the set of the indexes (J, L) such that ω · (L − J) = 0 reduces to the set J = L. Thus the resonant monomials are only the monomials of the form

ξ J η J = (ξ1 η1 )J1 ...(ξn ηn )Jn ≡ I1J1 ...InJn .

(31)

It follows that in the nonresonant case a function Z is in normal form if and only if it is a function of the actions only.

2.2.3 Proof of Birkhoff’s theorem We proceed by induction. The theorem is trivially true for r = 0. Supposing it is true for r we prove it for r + 1. First consider the Taylor polynomial of degree r + 3 (r) of R (r) and denote it by Hr+1 ∈ Hr+1 . Let χr+1 ∈ Hr+1 be the solution of the homological equation (r) {χr+1 ; H0 } + Hr+1 = Zr+1 (32) with Zr+1 in normal form. By Lemma 2.3 such a χr+1 exists. By corollary 6.1 below, χr+1 generates an analytic flow. Use it to generate the Lie transform φr+1 and consider H (r+1) := H (r) ◦ φr+1 and write it as follows

A Birkhoff normal form theorem for some semilinear PDEs

221

H (r) ◦ φr+1 = H0 + Z (r) +Zr+1 +(Z (r) ◦ φr+1 − Z (r) ) +H0 ◦ φr+1 − (H0 + {χr+1 ; H0 }) (r)

(33) (34) (35) (36)

+(R (r) − Hr+1 ) ◦ φr+1

(37)

(r) (r) +Hr+1 ◦ φr+1 − Hr+1

(38)

.

define Z (r+1) := Z (r) + Zr+1 . To prove that the terms (35–38) have a vector field with a zero of order at least r + 3 use Lemma 2.1 which ensures that each line is the remainder of a Taylor expansion (in the space variables) truncated at order r + 3. It remains to show that the estimate (9) of the deformation holds. Denote by Rr+1 (r) a positive number such that B2Rr+1 ⊂ Us , and remark that, by Lemma 6.2, possibly reducing Rr+1 , one has

φr+1 : Bρ → B2ρ ,

∀ρ ≤ Rr+1

and sup z − φr+1 (z) ≤ Cρ r+2 .

(39)

Bρ

Define Tr+1 := Tr ◦ φr+1 then one has Id − Tr+1 = Id − Tr ◦ φr+1 = Id − Tr + Tr − Tr ◦ φr+1 and thus, for any z ∈ Bρ with ρ small enough, we have z − Tr+1 (z) ≤ z − Tr (z) + Tr (z) − Tr (φr+1 (z)) ≤ Cr ρ 2 + sup dTr (z) sup z − φr+1 (z) z∈B2ρ

≤ Cr ρ +Cρ 2

r+2

z∈Bρ

≤ Cr+1 ρ 2

from which the thesis follows.

3 The case of PDEs 3.1 Hamiltonian formulation of the wave equation Consider the nonlinear wave equation (1). It is well known that the energy is a conserved quantity for (1). It is given by    2  u∆ u µ 2 u2 v − + dd x + F(u)dd x (1) H(u, v) := 2 2 2 D D

222

D. Bambusi

where v := ut and F is such that −F = f , and D is either Sd (d-dimensional sphere) or [0, π ]. The function H is also the Hamiltonian of the system and the corresponding Hamilton equations are given by u˙ = ∇v H ,

v˙ = −∇u H

(2)

where ∇u H is the L2 gradient of H with respect to u, defined by ∀h ∈ C∞ (D)

∇u H; hL2 = du Hh

(3)

where du is the differential with respect to the u variables. ∇v is defined similarly. To write (1) in the form (1) we have to introduce the basis of the eigenfunctions of the Laplacian. In the case of [0, π ] the eigenfunctions are given by 1 e1 := √ , π

e j := 

1 cos(( j − 1)x) , π /2

j≥2

(4)

and the corresponding eigenvalues of −∆ are λ j = ( j − 1)2 . In the case of the d dimensional sphere the eigenvalues λ j of −∆ are given by

λ j = ( j − 1)( j + d − 2) ;

(5)

moreover the jth eigenvalue has multiplicity   j+d −1 ∗ . l ( j) := d We will denote by e jl a basis of eigenfunctions of the Laplacian, which is orthonormal in L2 and such that −∆ e jl = λ j e jl ,

j≥1,

l = 1, ..., l ∗ ( j) .

For example they can be chosen to be the spherical harmonics. In both cases define ω j , p jl and q jl by  ω j := λ j + µ 2 q jl u = ∑ √ e jl , ωj jl

√ v = ∑ ω j p jl e jl

(6)

(7) (8)

jl

with the convention that l takes only the value 1 in the case of [0, π ] (and that, in such a case it will not be written). Then the Hamiltonian (1) takes the form (1) with H0 = ∑ ∑ ω j j

l

p2jl + q2jl 2

and HP is given by the second integral in (1) considered as a function of q jl .

(9)

A Birkhoff normal form theorem for some semilinear PDEs

223

3.2 Extension of Birkhoff’s theorem to PDEs: heuristic ideas In this section we will concentrate on the case of the nonlinear wave equation on [0, π ]. The main difficulty one meets in order to extend the theory of Birkhoff normal form to infinite dimensional systems rests in the denominators one meets in solving the homological equation, namely in the second of equations (26). Indeed, while in the finite dimensional case one has that the set of vectors with integer components having modulus smaller than a given r is finite, this is no longer true in infinite dimensions. It turns out that typically the denominators in (26) accumulate to zero already at  order 4. An example of such a behavior is the following one. Consider ω j+1 := j2 + µ 2 . For l ≥ 1 consider the integer vector K (l) whose only components different from zero are given by Kl = −2, Kl−1 = 1 Kl+1 = 1; such a vector has modulus 4, and one has K (l) · ω = ωl+1 + ωl−1 − 2ωl    µ2 = l 2 + µ 2 + (l − 2)2 + µ 2 − 2 (l − 1)2 + µ 2 ∼ 3 → 0 l Thus Birkhoff theorem does not trivially extend to infinite dimensional systems. However it turns out that in the case of PDEs the nonlinearity has a particular structure. As a consequence it turns out that most of the monomials appearing in the nonlinearity are small and do not need to be eliminated through the normalization procedure. To illustrate this behavior consider the map H s ([0, π ])  u → u2 ∈ H s ([0, π ]) ,

(10)

which is the first term of the nonlinearity of the nonlinear wave equation (1). The use of Leibniz formula together with interpolation inequality allows one to prove the so called Tame inequality, namely 0 20 0u 0 ≤ Cs u u . s 1 s

(11)

The key point is that, if u has only high frequency modes then its H 1 norm is much smaller than the H s norm. Indeed, assume that, for some large M one has u=

∑ uˆk ek

(12)

k≥M

then one has u21 =

∑ k2 |uˆk |2 = ∑

k≥M

k≥M

k2s k2(s−1)

|uˆk |2 ≤

1 u2s . M 2(s−1)

(13)

224

D. Bambusi

Collecting (13) and (11) one gets 0 20 0u 0 ≤ Cs 1 u2 , s s M s−1

(14)

which is very small if M and s are large. In order to exploit such a condition one can proceed as follows: given u ∈ H s split it into high frequency and low frequency terms, namely write uS :=

∑

uL :=

uˆk ek ,

|k| 0, and α ∈ R such that for any N large enough one has     γ   (17)  ∑ ω jKj ≥ α ,  N  j≥1 for any K ∈ Z∞ , fulfilling 0 = |K| := ∑ j |K j | ≤ r + 2, ∑ j>N |K j | ≤ 2.

228

D. Bambusi

It is easy to see that under this condition one can solve the Homological equation. The precise statement is given by the following lemma. Lemma 4.1. Let f be a homogeneous polynomial of degree less or equal than r having localized coefficients. Let H0 be given by (9) and assume that the frequency vector fulfills the condition r-NR. Consider the Homological equation {H0 , χ } + f = Z .

(18)

Its solution χ , Z defined by (26) has localized coefficients. In particular χ has localized coefficients.

4.2 Statement of the normal form theorem and its consequences Using the above results it is very easy to prove a version of the Birkhoff normal form theorem for PDEs. Definition 4.6. With reference to a system of the form (1) with H0 given by (9), the quantity p2jl + q2jl (19) J j := ∑ 2 l is called the total action of the modes with frequency ω j . Theorem 4.3. Fix r ≥ 1, assume that the nonlinearity HP has localized coefficients and that the frequencies fulfill the nonresonance condition (r-NR), then there exists (r) a finite sr a neighborhood Usr of the origin in Psr and a canonical transformation (r) T : Usr → Psr which puts the system in normal form up to order r + 3, namely H (r) := H ◦ T = H0 + Z (r) + R (r)

(20)

where Z (r) and R (r) have localized coefficients and (i) Z (r) is a polynomial

of degree r + 2 which Poisson commutes with J j for all j’s, namely J j ; Z (r) ≡ 0;

(ii) R (r) has a small vector field, i.e. 0 0 0X (r) (z)0 ≤ C zr+2 , sr R sr

∀z ∈ Usr ;

z − Tr (z)sr ≤ C z2sr ,

∀z ∈ Usr .

(r)

(21)

(iii) One has (r)

(22) −1 .

An inequality identical to (9) is fulfilled by the inverse transformation Tr (r) (r) (iv) For any s ≥ sr there exists a subset Us ⊂ Usr open in Ps such that the (r) restriction of the canonical transformation to Us is analytic also as a map from Ps → Ps and the inequalities (21) and (22) hold with s in place of sr . The proof is deferred to Sect. 6.2.

A Birkhoff normal form theorem for some semilinear PDEs

229

In order to deduce dynamical consequences we fix the number r of normalization steps; moreover, it is useful to distinguish between the original variables and the variables introduced by the normalizing transformation. So, we will denote by z = (p, q) the original variables and by z = (p , q ) the normalized variables, i.e. z = Tr (z ). More generally we will denote with a prime the quantities expressed in the normalized variables. Proposition 4.1. Under the same assumptions of Theorem 4.3, ∀s ≥ sr there exists ε∗s such that, if the initial datum fulfills

ε := z0 s < ε∗s then one has (i) z(t)s ≤ 4ε (ii)



for



∑ j2s J j (t) − J j (0) ≤ Cε M+3

|t| ≤

|t| ≤

for

j

and





∑ j2s J j (t) − J j (0) ≤ Cε 3

1 εr

for

(23) 1

ε r−M

|t| ≤

j

,

1 . εr

M 1 local existence is ensured only in H s , with s > 1, so that the energy norm is useless in order to deduce estimate of the existence times of solutions. At present the method of Birkhoff normal form is the only one allowing one to improve the times given by the local existence theory.

A Birkhoff normal form theorem for some semilinear PDEs

231

5 Discussion First I would like to mention that, as shown in [BG06], Theorem 4.3 is a theorem that allows one to deal with quite general semilinear equations in one space dimension. The limitation to semilinear equation is evident in Theorem 4.3. Thus in particular all the equations with nonlinearity involving derivatives are excluded from the present theory. It would be of major interest to have a theory valid also for some quasilinear equations, since most physical models have nonlinearities involving derivatives. Very little is known on quasilinear problems. At present the only known result is that of [DS04] (and a recent extension by Delort), where only one step of normal form was performed for the quasilinear wave equation. It would be very interesting to understand how to iterate the procedure developed in such papers. The limitation to one-space dimension is more hidden. Actually it is hidden in the nonresonance condition. Indeed its verification is based on the asymptotic behavior of the frequencies: the nonresonance condition is typically satisfied only if the frequencies grow at infinity at least as ω j ∼ j. As it was shown in the example of the nonlinear wave equation on the sphere, the possible multiplicity of the frequencies is not a problem. The theory easily extends to the case where the differences between couples of frequencies accumulate only at a discrete subset of R. The understanding of the structure of the frequencies in higher dimension is surely a key point for the extension of the theory to higher dimensions. Finally I would like to mention the fact that all known applications of the theory we are considering pertain to equations on compact manifolds, however in principle the theory applies to smooth perturbations of linear system with discrete spectrum. A nice example of such a kind of systems is the Gross Pitaevskii equation. It would be interesting to show that such an equation fulfills the assumption of Theorem 4.3. This could be interesting also in connection with the study of the blow up phenomenon.

6 Proofs 6.1 Proof of the properties of functions with localized coefficients Lemma 6.1. Let z ∈ Ps with s > ν + 1/2 then there exists a constant Cs such that 0 0 (1) ∑ | j|ν 0Π j z0 ≤ Cs zs j =0

Proof. One has

0 0 2 2 0 0 0 0 0 02 1 s Π jz 0 ∑ | j| Π j z ≤ ∑ | j| | j|s−ν ≤ ∑ | j|2(s−ν ) ∑ | j|2s 0Π j z0 j j j j =0 ν0

which is the thesis.



232

D. Bambusi

Proof of Theorem 4.1. Write explicitly the norm of F(z). One has F(z)2s

0 0 = ∑ |l| 0 0j l 2s 0

1

02 0 ˜ Π j1 z, ..., Π jr z)0 0 . ∑ Πl F( 0 ,..., jr

(2)

In what follows, to simplify the notation we will write 0 0 a j := 0Π j z0 . One has

0 0 0 0 0j

0 0 0 ˜ |l| Π Π z, ..., Π z) F( j j l ∑ r 0≤C 1 0 ,..., jr j s

1

∑

|l|s

1 ,..., jr

µ ( j, l)ν +N a j1 ...a jr S( j, l)N

(3)

Since this expression is symmetric in j1 , ... jr the r.h.s. of (3) is estimated by a constant times the sum restricted to ordered multi-indexes, namely indexes such that | j1 | ≤ | j2 | ≤ ... ≤ | jr |. Moreover, in order to simplify the notations we will restrict to the case of positive indexes. To estimate (3) remark that for ordered multi-indexes one has µ ( j, l) ≤ 2 jr . (4) l S( j, l) Indeed, if l ≤ 2 jr this is obvious (µ /S < 1 by the very definition), while, if l > 2 jr one has S( j, l) ≥ |l − jr | > l/2, and therefore l

µ ( j, l) ≤ µ ( j, l) ≤ 2 jr . S( j, l)

Remark now that, by the definition of S one has $ 1 + | jr − l| if l ≥ jr−1 S( j, l) ≥ µ ( j, l) + jr − jr−1 ≥ l + jr − jr−1 if l < jr−1 ˆ j, l) := min{1+| jr −l|, l + jr − jr−1 } and remark that S( j, l) ≥ S( ˆ j, l). Thus define S( Remark also that µ ( j, l) ≤ jr−1 . So it follows that (3) is smaller than (a constant times)

∑

j1 ,..., jr

µ ( j, l)N +ν jrs ˆ j, l)N S(

a j1 ...a jr ≤

∑

jrs

j1 ,..., jr

N +ν jr−1

ˆ j, l)N S(

a j1 ...a jr

(5)

≤ zr−2 s1

∑

jr−1 , jr

jrs

N +ν jr−1

ˆ j, l)N S(

a jr−1 a jr

(6)

where we denoted N := N − s and we used Lemma 6.1; we denoted by s1 a number such that s1 > 1/2. Inserting in (2) one gets

A Birkhoff normal form theorem for some semilinear PDEs

 F(z)2s

≤

=

zs2(r−2) 1 zs2(r−2) 1

∑ l

∑

jrs

jr−1 , jr



∑ ∑ l

⎛

N +ν jr−1

ˆ j, l)N S(

N +ν jr−1 a jr−1

jr−1

a jr−1 a jr

∑

jrs

jr

N +ν

≤ zs2(r−2) ∑ ⎝ ∑ jr−1 a jr−1 1 jr−1

l

233

2

a jr 1

/2 ˆ

N ˆ S( j, l) S( j, l)N /2

7 8 8 9

a2jr

2

∑ jr2s S(ˆ j, l)N jr

2

⎞2 1 ⎠ ∑ S(ˆ j, l)N jr

ˆ j, l) ≥ S( ˇ jr , l) := Now the last sum in jr is finite provided N > 1. Remark now that S( min{1 + |l − jr |, l} (independent of jr−1 ), and therefore the above quantity is estimated by a constant times 7 ⎛ ⎞2 8 2 8 a jr N +ν ⎠ zs2(r−2) a jr−1 9∑ jr2s (7) ∑ ⎝ ∑ jr−1

1 ˇ S( j, l)N j j l r

r−1

= zs2(r−2) 1

∑ jr

jr2s a2jr



1 ∑ S(ˇ j, l)N l

2

∑

N +ν jr−1 a jr−1

(8)

jr−1

z2s0 ≤ C z2s z2(r−2) s1

(9)

where s0 is such that s0 > N + ν + 1/2. Choosing s1 ≤ s0 and estimating zs1 with zs0 one gets the thesis.

 Proof of Theorem 4.2. First remark that the multilinear form associated to the polynomial d f (z)G(z) is given by the symmetrization of # (r1 ) , ..., z(r1 +r2 −1) )) . r1 f#(z(1) , ..., z(r1 −1) , G(z

(10)

We will estimate the coefficients of this multilinear function. This will give the result. Forgetting the irrelevant constant r1 , the quantity to be estimated is # Πi z, ..., Πi z)) f#(Π j1 z, ..., Π jr1 −1 z, G( r2 1 # Πi z, ..., Πi z)) = ∑ f#(Π j1 z, ..., Π jr1 −1 z, Πl G( r2 1 l

≤ CN,N ∑ l

0 0 0 µ ν1 +N ( j, l) µ ν2 +N (i, l) 0 0 0Π j z0 ... 0 Π z 0 0 i

r 1 N N 2 S( j, l) S(i, l)

(11) (12) (13)

Thus it is enough to estimate

µ ν1 +N ( j, l) µ ν2 +N (i, l) ∑ S( j, l)N S(i, l)N l This is the heart of the proof.

(14)

234

D. Bambusi

In order to simplify the notation we will restrict to the case r1 − 1 = r2 = r. Due to the symmetry of this estimate we will restrict the case of ordered indexes, that can also be assumed to be positive, so that one has jr ≥ jr−1 ≥ ... ≥ j1 and similarly for i. All along this proof we will use the notation ˜ j) := jr − jr−1 ≡ S( j) − µ ( j) S( We have to distinguish two cases. First case jr ≥ ir ≥ jr−1 . The proof of this first case is (up to minor changes) equal to that given in [Gré06]. Take N = N, then before estimating (14), we need to estimate the general term of the sum. So we collect a few facts on it. The main relation we need is ˜ j) ≤ S(i, ˜ l) + S( ˜ j, l) . S(i,

(15)

This will be established by writing explicitly all the involved quantities as l varies. ˜ j) = jr − ir . Then one has So, first remark that S(i, $ $ ir − ir−1 if l ≤ ir−1 j − jr−1 if l ≤ jr−1 ˜ ˜ S(i, l) = , S( j, l) = r |ir − l| if l > ir−1 | jr − l| if l > jr−1 which gives ⎧ ir − ir−1 + jr − jr−1 ≥ jr − jr−1 ≥ jr − ir if l ≤ ir−1 ⎪ ⎪ ⎪ ⎪ ⎨ l − ir + jr − jr−1 ≥ jr − jr−1 ≥ jr − ir if ir−1 < l ≤ jr−1 ˜ l) + S( ˜ j, l) = |ir − l| + | jr − l| ≥ | jk − l| ≥ jk − ik if jk−1 < l ≤ ir S(i, ⎪ ⎪ |i − l| + | j − l| = j − l + l − i if ir < l ≤ jr ⎪ r r r r ⎪ ⎩ |ir − l| + | jr − l| ≥ l − ir ≥ jr − ir if jr ≤ l from this (15) follows. One also has

µ ( j, l) ≤ µ (i, j) ,

µ (i, l) ≤ µ (i, j) .

(16)

Thus ˜ j) ˜ l)+S( ˜ j, l) ˜ l) S(l, ˜ j) S(i, l) S(l, j) S(i, S(i, S(i, S(i, j) = 1+ ≤ 1+ ≤ 1+ + < + µ (i, j) µ (i, j) µ (i, j) µ (i, l) µ (l, j) µ (i, l) µ (l, j) From this one has

$ % µ (i, j) 1 µ (i, l) µ (l, j) ≥ min , . S(i, j) 2 S(i, l) S(l, j)

Separate the sum over those l such that Let L1 be the first set. Then one has

µ (i,l) S(i,l)

>

µ (l, j) S(l, j)

(17)

and that over its complement.

A Birkhoff normal form theorem for some semilinear PDEs

∑

l∈L1

235

N−1−ε µ (i, l)1+ε +ν2 µ ν1 +N ( j, l) µ ν2 +N (i, l) N−1−ε ν1 µ (i, j) ≤ 2 µ ( j, l) ∑ S( j, l)N S(i, l)N S(i, j)N−1−ε S(i, l)1+ε l≥1

≤C

µ (i, j)N+ν1 +ν2 . S(i, j)N−1−ε

Acting in the same way for the case of L1c one concludes the proof in the first case. Second case jr ≥ jr−1 > ir . Here it is easy to see that (15) still holds. However, in some cases it happens that the equation

µ ( j, l) ≤ µ (i, j)

(18)

is violated. When (18) holds the proof of the first case extends also to the present case. So let us consider only the case where (18) is violated. We claim that in this case one has µ ( j, l) ≤ 2µ (i, j) . (19) ˜ l) S(i, To prove (19) we distinguish two cases (i) jr−2 ≤ ir ≤ jr−1 ≤ jr . Then (18) is violated when ir < l ≤ jr−1 . In this case one has ˜ l) = l − ir S(i, It follows that

(20)

l 1 µ ( j, l) 1 = ˜ S(i, l) µ (i, j) l − ir ir

which is easily seen to be smaller than 2 (for example write l = ir + δ , then the relation becomes evident). (ii) ir < jr−2 ≤ jr−1 ≤ jr . Here (18) is violated when jr−2 < l ≤ jr−1 . It is easy to see that also in this case (20) holds. Then l 1 l 1 µ ( j, l) 1 = ≤ ˜ l − ir i r S(i, l) µ (i, j) l − ir jr from which (19) still follows. It is now easy to conclude the proof. Take N = 2N + ν2 , then, using (19) one has

µ (i, l)ν1 +2N+ν2 µ ( j, l)N+ν2 µ (i, l)ν1 +2N+ν2 ≤ 2N+ ν N 2 S(i, l) S( j, l) S(i, l)N ≤



µ ( j, l) S(i, l)

N+ν2

µ (i, l)ν1 +2N+ν2 µ (i, j)N+ν2 S(i, l)N S( j, l)N

1 S( j, l)N

236

D. Bambusi

From this, following the proof given in the first case it is easy to prove that S(i, l) S( j, l) S(i, j) ≤ + µ (i, j) µ (i, l) µ (i, j) and to conclude the proof in the same way as in the first case.

 Proof of Lemma 4.1. Consider the polynomial f and expand it in Taylor series. Introduce now the complex variables (22). Remark that this is a linear change of variable so it does not change the degree of a polynomial. Remark that the change of variables does not mix the different spaces E j × E j . It follows that if a polynomial has localized coefficients in terms of the real variables p, q it has also localized coefficients when written in terms of the complex variables, i.e. it fulfills (10) with z j which is either ξ j or η j . Remark that the converse is also true. Now, Z is the sum of some of the coefficients of f so it is clear that its coefficients are still localized. In order to estimate χ , remark first that, in the particular case where f (z) ≡ f#(Π j1 ξ , ..., Π jr1 ξ , Πl1 η , ..., Πlr2 η ) (no summation over j, l) one has {H0 , f } = i(ω j1 + ... + ω jr1 − ωl1 − ... − ωlr2 ) f

(21)

It follows that in the case of general f the function χ solving the homological equation can be rewritten as

χ (ξ , η ) := ∑ jl

f#(Π j1 ξ , ..., Π jr1 ξ , Πl1 η , ..., Πlr2 η )

i(ω j1 + ... + ω jr1 − ωl1 − ... − ωlr2 )

(22)

where the sum runs over the indexes such that the denominators do not vanish. Now, it is easy to verify that by condition (r-NR) the denominators are bounded from below by γ /µ ( j, l)α . So χ fulfills the estimate (10) with ν substituted by ν + α , if  f does with ν .

6.2 Proof of the Birkhoff normal form Theorem 4.3 and of its dynamical consequences In this section we will fix s large enough and work in Ps . Here BR ⊂ Ps will denote the open ball of radius R with center at the origin in Ps . Moreover all along this section H j will denote the set of homogeneous polynomials of degree j + 2 having a Hamiltonian vector field which is smooth as map from Ps to itself. Finally, along this section we will omit the index s from the norm, thus we will simply denote . := .s .

A Birkhoff normal form theorem for some semilinear PDEs

237

First we estimate the domain where the Lie transform generated by a polynomial χ ∈ H j , ( j ≥ 1) is well defined. Lemma 6.2. Let χ ∈ H j , ( j ≥ 1) be a polynomial. Denote by φ t the flow of the corresponding vector field. Denote also   t¯ = t¯(R, δ ) := inf sup t > 0 : φ t (z) ∈ BR+δ and φ −t (z) ∈ BR+δ z∈BR

(minimum escape time of φ t (z) from BR+δ ). Then one has

δ 0 t¯ ≥ 0 2 0Xχ 0 R j+1

(23)

0 0 where 0Xχ 0 is the norm defined in remark 2.3. Moreover for any t, such that |t| ≤ t¯ and any z ∈ BR one has 0 0 0 0 t 0φ (z) − z0 ≤ |t|R j+1 0Xχ 0 (24) Proof. 0 First 0remark that, by the definition of t¯ one has that there exists z¯ ∈ BR such that 0φ ±t¯(¯z)0 = R + δ . Assume by contradiction t¯ < 2 X δ R j+1 , then, since for any  χ t with |t| < t¯ one has φ t (¯z) ∈ BR+δ . It follows that 0 t¯ 0 0 0 0 0 0 0 d s 0 t¯ 0 0 t¯ 0 0 φ (¯z)ds0 0φ (¯z)0 ≤ ¯z + 0φ (¯z) − z¯0 = ¯z + 0 0 0 ds  t¯ 0 0 0 0 ≤ R + 0Xχ (φ s (¯z))ds0 ≤ R + |t¯|R j+1 0Xχ 0 , 0

from which R + δ ≤ R + δ /2 which is absurd.

 Since χ is analytic together with its vector field (it is a smooth polynomial), then one has the following corollary. Corollary 6.1. Fix arbitrary R and δ , then the map

φ : Bσ × BR → BR+δ ,

δ 0 σ := 0 0 2 Xχ R j+1 0

(t, z) → φ t (z) is analytic. Here, by abuse of notation, we denoted by Bσ also the ball of radius σ contained in C. Proof of Theorem 4.3. The proof proceeds as in the finite dimensional case. The only fact that has to be ensured is that at any step the functions involved in the construction have localized coefficients. By Lemma 4.1 the solution χr+1 of the homological equation (32) has localized coefficients. Thus, by Theorem 4.1 its vector field is smooth on a space Psr+1 . This determines the index sr+1 of the space with minimal smoothness in which the transformation Tr+1 is defined. By corollary 6.1

238

D. Bambusi

χr+1 generates an analytic flow. As in the finite dimensional case we use it to generate the Lie transform. Then H (r+1) is still given by (35–38). Remark now that given a Hamiltonian function f , the Hamiltonian vector field of f ◦ φr+1 is given by −1 (φr+1 (z))X f (φr+1 (z)) X f ◦φr+1 (z) = dφr+1

(25)

so that the Hamiltonian vector fields of the terms (34), (35), (37), (38) are smooth. To ensure the smoothness of the vector field of (36) write (z) := H0 ◦ φ − H0 − {χr+1 , H0 } and remark that H0 (φr+1 (z)) − H0 (z) =

 1 d 0

 1

= 0

dt

t H0 (φr+1 (z))dt =

 1 0

t {χr+1 , H0 } (φr+1 (z))dt

(r)

t t (Hr+1 (φr+1 (z)) − Zr+1 (φr+1 (z))dt ,

where we used the homological equation to calculate {χr+1 , H0 }. Denote again G := (r) Hr+1 − Zr+1 , then one has  1

(z) = 0

t (G(φr+1 (z)) − G(z))dt ,

from which the smoothness of the vector field of (36) immediately follows. Since the Taylor expansion of the terms (35–38) can be computed using (13), by corollary 14 one has that all these functions have localized coefficients. Then, as in the finite dimensional case the terms (35–38) have a vector field with a zero of order at least r + 3 which ensures the estimate of the remainder. We show now that the normal form Z (r) commutes with all the J j . To this end remark that, by construction, the normal form contains only resonant monomials, i.e. monomials ξ L η J with   0 = ∑ ω j (J jl − L jl ) = ∑ ω j j

jl

∑(J jl − L jl )

.

(26)

l

Now the nonresonance condition implies  

∑(J jl − L jl )

∀j .

=0

l

It follows



Jj, ξ η

L J



 =i



∑(J jl − L jl ) l

which is the desired property.

ξ Lη J = 0

(27)

A Birkhoff normal form theorem for some semilinear PDEs

239

Finally the estimate (22) of the deformation can be obtained exactly as in the finite dimensional case.

 (r)

Proof of Proposition 4.1. We start by (i). Assume that ε is so small that B3ε ⊂ Us ; perform the normalizing transformation. Remark that, by (22), one has z 0 ∈ B2ε ⊂ (r) Us . Define F(z) := ∑ j | j|2s J j ≡ z2s , then, as far as z (t)s ≤ 3ε one has     t (r)    F(z (t)) − F(z 0 ) =   0 H , F (z (s))ds  t  

  (28) ≤  R (r) , F (z (s)) ds ≤ |t|Cε r+3 ≤ Cε 3 0

where the last inequality holds for the times (23). To conclude the proof of (23) it is enough to show that, for the considered times one actually has z (t) ∈ B3ε . To this end we follow the scheme of the proof of Lyapunov’s theorem: define 0 0 0 0   t¯ := sup t > 0 : 0z (t)0s < 3ε and 0z (−t)0s < 3ε To fix ideas assume that the equality is realized for t = t¯ Assume by contradiction that t¯ < ε −r , then one can use (28) which gives   0 02 0z (t¯)0 = 9ε 2 = F(t¯) ≤ F(z 0 ) + F(z (t)) − F(z 0 ) ≤ 4ε 2 +Cε 3 ,

(29)

which is impossible for ε small enough. We come to (ii). First remark that   (r) (r)

∂R

∂R ˙ + q jl , J j = ∑ −p jl ∂ q jl ∂ p jl l so that

   (r) (r)    ∂ R ∂ R   ∑ j2s J˙ j  = ∑ j2s −p jl ∂ q + q jl ∂ p  jl jl j jl ⎛ 2  2 ⎞⎞1/2 1/2 ⎛    (r)  (r)  ⎝∑ j2s ⎝ ∂ R  +  ∂ R  ⎠⎠ ≤ ∑ j2s (p jl 2 + q jl 2 )

 ∂ q jl   ∂ p jl  jl jl 0 0 0 0 0 0r+3 ≤ 0z 0s 0XR (r) (z )0s ≤ C 0z 0s

(30)

(31) (32)

which implies (24). To prove (25) write         J j (t) − J j (0) ≤ J j (z(t)) − J j (z (t)) + J (t) − J (0) + J j (z0 ) − J j (z 0 ) . (33)

240

D. Bambusi

The contribution of the middle term is estimated by (24). To estimate the contribution of the first and the last term write     (34) j2s q2jl − q jl 2  ≤ j2s 2|q jl | |q jl − q jl | + |q jl − q jl |2 adding the corresponding estimate for the p variables and summing over jl one gets the thesis. We come to (iii). In the considered case J j reduces to I j , so the actions are individually conserved. In this proof we omit the index l which would take only the p 2 (0)+q j 2 (0) value 1. Denote I¯j := j and define the torus 2

  T 0 := z ∈ Ps : I j (z ) = I¯j One has



d(z

(t), T 0 ) ≤

   2 1/2  ∑ j  I j (t) − I¯j  2s 

(35)

j

Notice that for a, b ≥ 0 one has, √ √     a − b ≤ |a − b| . Thus, using (32), one has that  2 d(z (t), T 0 ) ≤ ∑ j2s |I j (t) − I¯j | ≤ Cε M+3 j

Define now T0 := Tr (T 0 ) then, since Tr is Lipschitz one has d(z(t), T0 ) = d(Tr (z (t)), Tr (T 0 )) ≤ Cd(z (t), T 0 ) ≤ Cε

M+3 2

.



6.3 Proof of Proposition 4.2 on the verification of the property of localization of coefficients In this subsection we will prove the property of localization of coefficients for the  function u → u3 in the case where the basis used for the definition (10) is the basis of the eigenfunction of general second order elliptic operator. Thus the present theory directly applies also to the case of the equation utt − uxx +Vu = f (x, u) with Neumann boundary conditions on [0, π ]. The case of Dirichlet boundary conditions can also be covered by a minor variant (indeed in such a case the function   u → u3 has to be substituted by the function u → u4 .

A Birkhoff normal form theorem for some semilinear PDEs

241

Thus consider a second order elliptic operator P, which is L2 self adjoint. This means that we assume that in any coordinate system there exist smooth functions Vα (x), α ∈ Nd such that P = ∑|α |≤2 Vα ∂ α , where we used a vector notation for the derivative. Moreover we will assume that us+2 ≤ Pus . Then, by L2 symmetry, one gets 0 0 0 0 us ≤ 0Ps/2 u0 , 0

(36)

(where Ps/2 is defined spectrally). We will denote by D(Pk ) the domain of Pk . Finally, denote by λn the sequence of the eigenvalues of P counted without multiplicity (i.e. in such a way that λn+1 > λn ). We will assume that the eigenvalues of P behave as λn ∼ n2 . We will denote by En the eigenspace of P relative to λn . Let A be a linear operator which maps D(Pk ) into itself for all k ≥ 0, and define the sequence of operators AN := [P, AN−1 ] ,

A0 := A .

(37)

Lemma 6.3. Let P be as above and let u j ∈ En j . Then, for any N ≥ 0 one has |Au1 ; u2 | ≤

1 |AN u1 ; u2 | |λn1 − λn2 |N

(38)

Proof. One has A1 u1 ; u2  = [A, P]u1 ; u2  = APu1 ; u2  − PAu1 ; u2  = λn1 Au1 ; u2  − Au1 ; Pu2  = (λn1 − λn2 )Au1 ; u2  Equation (38) follows applying the above equality to the operator AN := [P, AN−1 ] and using an induction argument.

 To conclude the proof we have to estimate the matrix elements of AN , i.e. the r.h.s. of (38). To this end we need a few remarks and lemma. Remark 6.1. Consider two d-dimensional multi-indexes α and β and define   α! α , := β β !(α − β )! with the convention that it is 0 if β j > α j for some j. One has   α ∂ α (uv) = ∑ ∂ β u∂ α −β v . β β

(39)

242

D. Bambusi

Remark 6.2. Let A := a(x)∂ α and B := b(x)∂ β with a and b smooth functions. Then one has       β α γ γ a ∂ b−b ∂ a ∂ α +β −γ . (40) [A, B] = ∑ γ γ γ ≤α +β j

j

j

Lemma 6.4. Choose a coordinate system, let A = a0 (x) be a multiplication operator, then one has AN =

∑

|α |≤N

cα ∂ α

(41)

Vαβ (x)∂ β a0

(42)

(N)

(N)

with cα of the form (N)

cα =

∑

(N)

|β |≤2N−|α |

and Vαβ which are C∞ functions depending only on the functions Vα defining the operator P. (N)

Proof. First remark that by (40), the operator AN is a differential operator of order N. By induction, using (40) one easily sees that the coefficients of such an operator are linear combinations of the derivatives of a0 . To show (42) we proceed by induction. The result is true for N = 0. Then use equation (40) to compute ! " (N) Vα ∂ α ; cβ ∂ β =

∑

γ j ≤α j +β j

      β (N) (N) α Vα ∂ γ cβ − cβ ∂ γ Vα ∂ α +β −γ γ γ

(43)

Consider the first term in the square bracket which is the one involving more deriv(N) (N) (N) atives of cβ . Since cβ depends on ∂ δ a0 with |δ | ≤ 2N − |β |, one has that ∂ γ cβ

depends only on the derivatives ∂ δ a0 with |δ | ≤ 2N − |β | + |γ |; in order to conclude the proof we have to show that this is smaller than 2(N + 1) − (|α | + |β | − |γ |), a  fact which is true since |α | ≤ 2.

Remark 6.3. Let un ∈ En then by (36) one has un s ≤ Cns un 0

Remark 6.4. Let un ∈ En with un 0 = 1, and bα be a smooth function (α ∈ Nd ), then one has for any ν0 > d/2 one has bα ∂ α un 0 ≤ Cν0 bα ν0 n|α |

(44)

Remark 6.5. Let un ∈ En with un 0 = 1, and let bα := Vαβ (x)∂ β un (N)

(45)

A Birkhoff normal form theorem for some semilinear PDEs

243

(with some β ) then one has bα ν0 ≤ Cnν0 +|β |

(46)

(N)

with a C that depends on Vαβ . End of the proof of Proposition 4.2. Assume that n3 ≤ n2 ≤ n1 so that µ (n) = n3 and S(n) = n3 + n1 − n2 . Write the l.h.s. of (29) as |Aun2 ; un1 |

(47)

with A the multiplication operator by un3 . Using (38) this is smaller than 1 |n21 − n22 |N

AN un2 L2 un1 L2 .

(48)

To estimate AN un2 L2 we use (41) and estimate each term separately. By Sobolev embedding theorem, one term is estimated by 0 0 0 0 0 (N) 0 0 (N) α 0 0cα ∂ un2 0 ≤ C 0cα 0 ∂ α un2  ν0

ν0 > d/2. Using (42), (44), (46) one gets 0 0 0 0 0 0 (N) 0 2N+ν0 −|α | 0 0un 0 2 . 0cα 0 ≤ C 0un3 02N+ν −|α | ≤ Cn3 3 L ν0

0

where we used the ellipticity of P. We thus get that the l.h.s. of (29) is estimated by C

∑

|α |≤N

2N+ν0 −|α | |α | n2

n3

1 |n21 − n22 |N

0 0 un1 L2 un2 L2 0un3 0L2

A part from a constant, the sum of the coefficients in front of the norms is estimated by 2N+ν0

n3



n2 n3

N

1 = |n21 − n22 |N



n2 n1 + n2

N

ν +N

n30

|n1 − n2 |N

ν +N

≤

n30

|n1 − n2 |N

(49)

To conclude the proof just remark that n3 = µ , S = µ + (n1 − n2 ) and that if n3 > n1 − n2 then the inequality (29) is trivially true. On the contrary, if n3 ≤ (n1 − n2 ) the r.h.s. of (49) is smaller than ν +N

n30 which concludes the proof.



2 (n3 + |n1 − n2 |)N

244

D. Bambusi

6.4 Proof of Theorem 4.4 on the nonresonance condition The proof follows the proof of Theorem 6.5 of [Bam03] (see also [BG06]). We repeat the main steps for completeness. Fix r once for all and denote by C any constant depending only on r. The value of C can change from line to line. Finally we will denote m := µ 2 . Lemma 6.5. For any K ≤ N, consider K indexes j1 < ... < jK ≤ N; consider the determinant    ω j1 ω j2 . . . ω jK   dω d ω j2 dω j  . . . dmjK   dm1 dm   . . ... . D :=  (50)    . . ... .    d K−1 ω j d K−1 ω j d K−1 ω jK  1 2  ... dmK−1

One has

∏ l

Proof. One has

dmK−1





 D=C

dmK−1

ωi−2K+1 l

∏

(λ jl − λ jk )

≥

1≤l0 Jγ and remark that its complement is the union of a numerable infinity of sets of zero measure.

References [Bam99]

D. Bambusi. On long time stability in Hamiltonian perturbations of non-resonant linear PDEs. Nonlinearity, 12:823–850, 1999. [Bam03] D. Bambusi. Birkhoff normal form for some nonlinear PDEs. Comm. Math. Physics, 234:253–283, 2003. [BDGS07] D. Bambusi, J.M. Delort, B. Grébert, and J. Szeftel. Almost global existence for Hamiltonian semi-linear Klein-Gordon equations with small Cauchy data on Zoll manifolds. Comm. Pure Appl. Math., 2007. [BG06] D. Bambusi and B. Grébert. Birkhoff normal form for partial differential equations with tame modulus. Duke Math. J., 135(3):507–567, 2006.

A Birkhoff normal form theorem for some semilinear PDEs [BGG85]

[Bou96a]

[Bou96b] [Bou97] [Bou98] [Bou00] [Bou05]

[CW93] [DS04]

[DS06]

[EK06] [Gré06] [Kla83] [KP96] [Kuk87] [Way90] [XYQ97] [Yua06]

247

G. Benettin, L. Galgani, and A. Giorgilli. A proof of Nekhoroshev’s theorem for the stability times in nearly integrable Hamiltonian systems. Celestial Mech., 37:1–25, 1985. J. Bourgain. Construction of approximative and almost-periodic solutions of perturbed linear Schrödinger and wave equations. Geometric Functional Anal, 6:201– 230, 1996. Jean Bourgain. On the growth in time of higher Sobolev norms of smooth solutions of Hamiltonian PDE. Internat. Math. Res. Notices, (6):277–304, 1996. J. Bourgain. On growth in time of Sobolev norms of smooth solutions of nonlinear Schrödinger equations in RD . J. Anal. Math., 72:299–310, 1997. J. Bourgain. Quasi-periodic solutions of Hamiltonian perturbations of 2D linear Schrödinger equation. Ann. Math., 148:363–439, 1998. J. Bourgain. On diffusion in high-dimensional Hamiltonian systems and PDE. J. Anal. Math., 80:1–35, 2000. J. Bourgain. Green’s function estimates for lattice Schrödinger operators and applications, volume 158 of Annals of Mathematics Studies. Princeton University Press, Princeton, NJ, 2005. W. Craig and C. E. Wayne. Newton’s method and periodic solutions of nonlinear wave equations. Comm. Pure Appl. Math., 46:1409–1498, 1993. J. M. Delort and J. Szeftel. Long–time existence for small data nonlinear Klein– Gordon equations on tori and spheres. Internat. Math. Res. Notices, 37:1897–1966, 2004. J.-M. Delort and J. Szeftel. Long-time existence for semi-linear Klein-Gordon equations with small Cauchy data on Zoll manifolds. Amer. J. Math., 128(5):1187–1218, 2006. H. L. Eliasson and S. B. Kuksin. KAM for non-linear Schrödinger equation. Preprint, 2006. B. Grébert. Birkhoff normal form and Hamiltonian PDES. Preprint., 2006. S. Klainerman. On “almost global” solutions to quasilinear wave equations in three space dimensions. Comm. Pure Appl. Math., 36:325–344, 1983. S. B. Kuksin and J. Pöschel. Invariant Cantor manifolds of quasi-periodic oscillations for a nonlinear Schrödinger equation. Ann. Math., 143:149–179, 1996. S. B. Kuksin. Hamiltonian perturbations of infinite-dimensional linear systems with an imaginary spectrum. Funct. Anal. Appl., 21:192–205, 1987. C. E. Wayne. Periodic and quasi-periodic solutions of nonlinear wave equations via KAM theory. Comm. Math. Physics, 127:479–528, 1990. J. Xu, J. You, and Q. Qiu. Invariant tori for nearly integrable Hamiltonian systems with degeneracy. Math. Z., 226:375–387, 1997. X. Yuan. A KAM theorem with applications to partial differential equations of higher dimensions. Commun. Math. Phys to appear (Preprint 2006).

Normal form of holomorphic dynamical systems Laurent Stolovitch1

Abstract This article represents the expanded notes of my lectures at the ASI “Hamiltonian Dynamical Systems and applications”. We shall present various recent results about normal forms of germs of holomorphic vector fields at a fixed point in Cn . We shall explain how relevant it is for geometric as well as for dynamical purpose. We shall first give some examples and counter-examples about holomorphic conjugacy. Then, we shall state and prove a main result concerning the holomorphic conjugacy of a commutative family of germs of holomorphic vector fields. For this, we shall explain the role of diophantine condition and the notion of singular complete integrability.

1 Definitions and examples Let us consider the pendulum with normalized constants :

θ¨ + sin θ = 0

(1)

We would like to understand the behavior of the motion for the small oscillations of the pendulum, that is to say when θ is small. We are tempted to say that sin θ is well approximated by θ and then we would like to consider the much simpler equation (∗) θ¨ + θ = 0 instead of (1). If we set θ1 = θ and θ2 = θ˙ , equation (∗) can be written as & θ˙1 = θ2 . θ˙2 = −θ1

1 CNRS

UMR 5580, Laboratoire Emile Picard, Université Paul Sabatier, 118 route de Narbonne, 31062 Toulouse cedex 9, France e-mail: [email protected]

W. Craig (ed.), Hamiltonian Dynamical Systems and Applications, 249–284. c 2008 Springer Science + Business Media B.V.


249

250

L. Stolovitch

The dynamic is completely understood. Its trajectories are circles θ12 + θ22 = constant. Are these information relevant for the understanding of the dynamic of the original problem (1)? Does the closeness of equation (∗) to equation (1) imply that they have the same dynamical properties? In general, both answer are ‘No!’. In these lectures, we shall explain these phenomena and how to define a reasonable simplified problem to study : a normal form. Let us start with a very elementary example of a similar problem. In order to study the iterates of a square complex matrix A of Cn , that is the orbits {Ak x}k∈N for x ∈ Cn near the “fixed point” 0, it is very convenient to transform, with the help of a linear change of coordinates P, the matrix A into a Jordan matrix J = S + N, with S a diagonal matrix, N an upper triangular nilpotent matrix commuting with S : PAP−1 = S + N. Using the (block diagonal) structure of S + N, it is easy to study its iterates. Since Ak = P−1 J k P, we have An x = P−1 (J n y) where x = P−1 y. We thus obtain all informations needed for the study of the iterates of A. One of the great ideas of Poincaré was to try to proceed in the same way for vector fields. Is it possible to transform a given vector field X, vanishing at the origin of Rn (resp. Cn ), into a “simpler” one with the help of a local diffeomorphism Φ near the origin and which maps the origin to itself? The group of germs of Ck (resp. holomorphic, formal) diffeomorphisms at 0 ∈ Cn and tangent to IdCn at the origin, acts on the space of germs of holomorphic (resp. formal) vector fields at 0 ∈ Cn by conjugacy : if X is any representative of a germ of vector field X, and φ is any representative of a germ of diffeomorphism Φ , then Φ∗ X is the germ of vector field defined by φ∗ X(φ (x)) := Dφ (x)X(x) where Dφ (x) denotes the derivative of φ at the point x. One may first attempt to linearize formally X, that is to find a formal change of coordinates Φˆ , such that Φˆ ∗ X(y) = DX(0)y. Assume it is so then, one could expect to understand all about the dynamics of X since the flow of the linear vector field DX(0)y is easy to study. Nevertheless, this cannot be the case unless we are able to pull-back these informations by Φˆ , and this requires some “regularity” conditions on Φˆ . Is there a Ck (resp. smooth) linearizing diffeomorphism? When we are working in the analytic category, this regularity condition should be that Φˆ is holomorphic in a neighborhood of the origin. What happens in this situation? These ideas have been widely developed by V.I. Arnol’d and his school. Our main reference for this topic is the great book by V.I. Arnol’d [Arn88a]. We refer also to [AA88] which contain a lot of references on this topic. Singularities of mappings are also studied in the same spirit [AGZV85, AGZV88].

1.1 Vector fields and differential equations Let us consider a germ of vector field X at a point p : in a coordinate chart at p, it can be written X(z) = ∑ni=1 Xi (z) ∂∂zi . It is equivalent to consider the system of autonomous differential equations :

Normal form of holomorphic dynamical systems

251

⎧ z˙ = X1 (z) ⎪ ⎪ ⎨.1 .. . ⎪ ⎪ ⎩ z˙ = X (z) n

n

The Lie derivative of a germ of function f along the vector field X is the germ of function n ∂f (z). LX f (z) := ∑ Xi (z) ∂ zi i=1 It will also be denoted by X( f ). We will denote by [X,Y ] the Lie bracket of the vector fields X = ∑ni=1 Xi ∂∂zi and

Y = ∑ni=1 Yi ∂∂zi . It is defined to be n

[X,Y ] = ∑

i=1



∂ Yi ∂ Xi ∑ X j ∂ z j −Y j ∂ z j j=1 n



∂ . ∂ zi

It is skew-symmetric and satisfies the Jacobi identity : [X, [Y, Z]] + [Y, [Z, X]] + [Z, [X,Y ]] = 0. Moreover, if X,Y are vector fields and f a function [X, fY ] = f [X,Y ] + LX ( f )Y.

(2)

Two vector fields X,Y are said to be commuting pairwise whenever [X,Y ] ≡ 0. From the dynamical point of view, let us start at a point p, then let us follow the flow of X during a time t then follow the flow of Y during a time s. Let q be this end point. Let us start at p again but now follow the flow of Y during a time s first then follow the flow of X during a time t. Let q be this end point. The fact that X and Y commute pairwise means that q = q .

1.1.1 Notations Let us set some notations which will be used all along this article : let k ≥ 1 be an integer, • Pnk denotes the C-space of homogeneous polynomial vector fields on Cn and of degree k • Pnm,k denotes the C-space of polynomial vector fields on Cn , of order ≥ m and of degree ≤ k (m ≤ k) .k denotes the C-space of formal vector fields on Cn and of order ≥ k at 0 • X n • Xnk denotes the C-space of germs of holomorphic vector fields on (Cn , 0) and of order ≥ k at 0 • pkn denotes the C-space of homogeneous polynomial on Cn and of degree k

252

L. Stolovitch

.k denotes the C-space of formal power series on Cn and of order ≥ k at 0 • M n • Mnk denotes the C-space of germs of holomorphic functions on (Cn , 0) and of order ≥ k at 0 • O-n denotes the ring of formal power series in Cn • On denotes the ring of germs at 0 of holomorphic functions in Cn

1.1.2 Norms Let f ∈ C[[x1 , . . . , xn ]] be a formal power series : f = ∑Q∈Nn fQ xQ . We define f¯ as the formal power series f¯ = ∑Q∈Nn | fQ |xQ . We will say that a formal power series g dominates a formal power series f , if ∀Q ∈ Nn , | fQ | ≤ |gQ |. In that case, we will write f ≺ g. More generally, let q ≥ 1 be an integer and let F = ( f1 , . . . , fq ) and G = (g1 , . . . , gq ) be elements of (C[[x1 , . . . , xn ]])q ; we shall say that G dominates F, and we shall write F ≺ G, if fi ≺ gi for all 1 ≤ i ≤ q. We shall write F¯ = ( f¯1 , . . . , f¯q ). We shall say that F is of order ≥ m (resp. polynomial of degree ≤ m), if each of his components is of order ≥ m (resp. polynomial of degree ≤ m). Let r be an positive number and ( f , F, G) ∈ O-n × O-nq × O-nq , we define | f |r :=

∑ n | fQ |r|Q| = f¯(r, . . . , r)

Q∈N

and |G|r = maxi |gi |r ; these may not be finite. We have the following properties f G ≺ f¯G¯ if F ≺ G then |F|r ≤ |G|r ∂F ∂ F¯ = ∂ xk ∂ xk Let us define Hnq (r) = {F ∈ O-nq | |F|r < +∞}; |.|r is norm on this space. Together with the norm |.|r , this space is a Banach space (see [GR71]). Lemma 1.1.1 Let F = ∑Q∈Nn FQ xQ an element of Hnq (r), then we have the following inequalities : Fr ≤ |F|r  m R |F|r if ord(F) ≥ m, R ≤ r |F|R ≤ r d |DF|r ≤ |F|r if F is a polynomial of degree ≤ d r

(3) (4) (5)

Here Fr denotes the supremum of |F(z)| on the polydisc |zi | < r, 1 ≤ i ≤ l. Proof. The first inequality comes from the fact that for all x in the polydisc of radius r, we have

Normal form of holomorphic dynamical systems

253

     Q  ∑ FQ x  ≤ ∑ |FQ ||xQ | ≤ |F|r . Q∈Nn  Q∈Nn For the second, we have

∑ n

Q∈N , |Q|≥m

|FQ |R|Q| ≤ Q∈N

R|Q| Rm |Q| |F |r ≤ |FQ |r|Q| . Q |Q| rm Q∈Nn∑ , |Q|≥m r , |Q|≥m

∑ n

For the last one, we have F = ∑Q∈Nn , |Q|≤d FQ xQ . Hence, we have     ∂F  Q−E j  | |r =  F q x  Q j Q∈Nn∑  ∂xj , |Q|≤d

r

=

∑

|FQ |q j r|Q|−1

Q∈Nn , |Q|≤d

≤

d |F|r . r

We shall often use the estimate |(DG).F|r ≤ n|DG|r |F|r whenever (F, G) ∈ Hnn (r). Lemma 1.1.2 [Sto00][Prop. 3.1.1] Let r > 0, a ∈ C∗ and g ∈ Hn1 (r). We assume that |g|r < |a|. Then    1  1    a + g  ≤ |a| − |g|r r

1.2 Normal forms of vector fields In the sequel, we will assume that the linear part of X at the origin is not nilpotent (see [CS86] normal form with nilpotent linear part) and for the sake of simplicity we even assume that is it semi-simple : n

S := DX(0)x = ∑ λi xi i=1

∂ ∂ xi

is a nonzero diagonal vector field. If Q = (q1 , . . . , qn ) ∈ Nn , we will write (Q, λ ) := q q ∑ni=1 qi λi , |Q| := q1 + · · · + qn and xQ := x11 · · · xnn . Proposition 1.2.1 (Poincaré–Dulac normal form) Let X = S + R2 be a nonlinear perturbation of the linear vector field S. Then there exits a formal change of coordinates Φˆ tangent to the identity such that ˆ Φˆ ∗ X = S + N, ˆ = 0. where the nonlinear formal vector field Nˆ commutes with S : [S, N]

254

L. Stolovitch

By formal change of coordinates Φˆ tangent to the identity, we mean that there exists formal power series φˆi (x) = ∑Q∈Nn ,|Q|≥2 φi,Q xQ ∈ C[[x1 , . . . xn ]] of order ≥ 2, such that Φˆ i (x) = xi + φˆi (x), the ith-component of Φˆ . Let us describe a normal form in local coordinates. First of all, we notice that   ∂ ∂ = ((Q, λ ) − λi ) xQ . S, xQ ∂ xi ∂ xi Therefore, such an elementary vector field commute with S if and only if (Q, λ ) = λi . This is called a resonance relation and xQ ∂∂xi the associated resonant vector field. Therefore, the formal normal form proposition can be rephrased as : there exists a formal diffeomorphism Φˆ (which is not unique in general) such that   n n ∂ ∂ Φˆ ∗ X = ∑ λi xi +∑ ai,Q xQ ∑ ∂ x ∂ xi i i=1 i=1 (Q,λ )=λ i

where the sum is over the multiintegers Q ∈ Nn , |Q| ≥ 2 and the index i which satisfy to (Q, λ ) = λi and where the ai,Q ’s are complex numbers. Example 1.2.2 Let ζ be a positive irrational number. Let us consider the vector field X & x˙ = x + f (x, y) y˙ = −ζ y + g(x, y) where f , g are smooth functions vanishing at the origin as well as their first derivatives. It is formally linearizable since the only integer solution (q1 , q2 ) of q1 − ζ q2 = 0 is (0, 0). Hence, there are no resonance relation satisfied. Example 1.2.3 Let us consider the vector field X & x˙ = 2x + y2 + f (x, y) y˙ = y + g(x, y)

(6)

where the smooth functions f , g vanish at order 3 at the origin. There is one and only one resonance relation satisfied : 0λ1 + 2λ2 = λ1 . Therefore, X is formally conjugate to the normal form & x˙ = 2x + y2 . (7) y˙ = y Example 1.2.4 Let us consider the analytic vector field X & x˙ = x + f (x, y) y˙ = −y + g(x, y)

(8)

Normal form of holomorphic dynamical systems

255

for some holomorphic functions f , g vanishing at first order at the origin. It is clear that the only solutions of the resonance relation q1 λ1 + q2 λ2 = λ1 (resp. q1 λ1 + q2 λ2 = λ2 ) are of the form q1 = q2 + 1 (resp. q2 = q1 + 1). Thus, the resonant vector fields are generated by (xy)l x ∂∂x and (xy)l y ∂∂y where l is a positive integer. Applying Poincaré-Dulac theorem to equation (8) leads to a formal normal form & ˆ x˙ = xF(xy) (9) ˆ y˙ = −yG(xy) ˆ Gˆ are formal power series which values at 0 is 1. where F, Example 1.2.5 Let us extend example 1.2.4 by Example 1.2.3 in a four-dimensional system : w˙ = w + e(w, x, y, z) x˙ = −x + f (w, x, y, z) y˙ = 2iy + g(w, x, y, z) z˙ = iz + h(w, x, y, z) Its formal normal form is of the form ˆ w˙ = wF(wx) ˆ x˙ = −xG(wx) y˙ = 2iHˆ 1 (wz)y + Hˆ 2 (wz)z2 z˙ = iHˆ 3 (wz)z Example 1.2.6 Let us consider the five-dimensional system ⎧ x˙1 = x1 + f1 (x) ⎪ ⎪ ⎪ ⎪ ⎨ x˙2 = −x2 + f2 (x) x˙3 = −ζ x3 + f3 (x) ⎪ ⎪ x˙4 = ix4 + f4 (x) ⎪ ⎪ ⎩ x˙5 = ix5 + f5 (x) where ζ is a positive irrational number. Its normal form is of the form ⎧ x˙1 = x1 fˆ1 (x1 x2 ) ⎪ ⎪ ⎪ ⎪ ⎨ x˙2 = −x2 fˆ2 (x1 x2 ) x˙3 = −ζ fˆ3 (x1 x2 )x3 ⎪ ⎪ ⎪ x ˙ = ix4 + x4 gˆ1,1 (x1 x2 ) + x5 gˆ1,2 (x1 x2 ) ⎪ ⎩ 4 x˙5 = ix5 + x4 gˆ2,1 (x1 x2 ) + x5 gˆ2,2 (x1 x2 )

(10)

where the fˆi ’s (resp. gˆi, j ’s) are formal power series of one variable (resp. vanishing at the origin).

256

L. Stolovitch

1.2.1 Hamiltonian vector fields We refer the Arnold book [Arn97] for this section. To a germ of function H : (R2n , 0) → (R, 0) vanishing at first order at 0, we can associate a germ of vector field XH of (R2n , 0) vanishing at the origin. If (x, y) are local coordinates, it is defined to be

∂H , j = 1, . . . , n ∂yj ∂H y˙ j = − , j = 1, . . . , n. ∂xj x˙ j =

It is called the Hamiltonian vector field associated to H. The function H is called the Hamiltonian of XH . Definition 1.2.7 A change of coordinate X j = φ j (x, y), Y j = ψ j (x, y) is called canonical if it preserve the symplectic form ω = ∑nj=1 dx j ∧ dy j . In other words, n

n

j=1

j=1

∑ dx j ∧ dy j = ∑ dX j ∧ dY j .

If we conjugate an Hamiltonian vector field XH by a canonical diffeomorphism Φ , we obtain again an Hamiltonian vector field, namely XH◦Φ . We shall say that the Hamiltonian H is a Birkhoff normal form whenever its associated Hamiltonian vector field XH is a normal form. Definition 1.2.8 In symplectic coordinates (x, y), we define the Poisson bracket of the germ of functions to be  n  ∂ f ∂g ∂g ∂ f . − { f , g} := ∑ ∂xj ∂yj j=1 ∂ x j ∂ y j It satisfies the following properties : • {., .} is bilinear and skew-symmetric • { f , {g, h}} + {g, {h, f }} + {h, { f , g}} = 0 (Jacobi identity) • { f , gh} = { f , g}h + { f , h}g (Leibniz identity) It is easy to show that [XH , XG ] = X{H,G} .

1.3 Examples about linearization Example 1.3.1 The normal form (7) is topologically conjugate to the linear part & x˙ = 2x . y˙ = y

Normal form of holomorphic dynamical systems

257

By this, we mean there exists an homeomorphism H fixing the origin which maps the trajectories of the normal form to the trajectories of its linearized : H(φt (x)) = φtlinearized (H(x)) where φt denotes the flow at time t starting at x. This is a consequence of Hartman–Grobmann theorem. Nevertheless, it can be shown that the normal form is not C2 -conjugate to its linearized at the origin. Theorem 1.3.2 [Ste58, Bru95] Assume the linear part S is non-resonant, i.e. there is no resonance relation satisfied. Then any smooth nonlinear perturbation X = S + R of S is smoothly conjugate to its linear part S. What happens in the analytic context? Example 1.3.3 We borrow this example to J.-P. Françoise [Fra95]. Let us consider a special case of Example 1.2.2. Let us assume that the irrational number ζ is Liouvillian. By this, we mean that there exists two sequences of positive integers (pn ), (qn ) both tending to infinity with n such that     1 ζ − pn  < .  qn  qn (qn !) The number ζ is too well approximated by rational numbers. Given such a pair of sequence, let us consider the function f (x, y) =

1 . 1 − ∑ x pn yqn

It is holomorphic in a neighborhood of the origin and f (0) = 1. Let us set S := x ∂∂x − ζ y ∂∂y and let us consider the germ of holomorphic vector field defined to be X = f (x, y)S. Its linear part at the origin is S. Let us find the formal change of coordinate that linearizes it (in this case, it’s unique) : x˜ = x exp(−V (x, y)), y˜ = y exp(−W (x, y)). Then, ˜ = xLX (−V (x, y)) exp(−V (x, y)) x exp(−V (x, y)) = x˜ = LX (x) + exp(−V (x, y))LX (x). Here, the first equality comes from the definition, the second comes from that fact that X is linearized in the new coordinates. Therefore, we have that LX (V ) = f − 1 which is equivalent to f −1 = ∑ x pn yqn . LS (V ) = f This equation has the unique solution V =∑ which is divergent at the origin since

1 x pn yqn pn − ζ qn 1 pn −ζ qn

≥ qn !.

258

L. Stolovitch

This example shows that one need an “arithmetical” condition on the small divisors (Q, λ ) − λi = 0. The major step in the understanding of the phenomenon is due to C.L Siegel. Definition 1.3.4 We shall say that λ = (λ1 , . . . , λn ) is diophantine of type ν ≥ 0 if there exists C > 0 such that, for all multiindexes Q ∈ Nn , |Q| ≥ 2, |(Q, λ ) − λi | >

C . |Q|µ

We shall say that there no small divisor if there exists a constant c > such that |(Q, λ ) − λi | > c. Theorem 1.3.5 [Sie42] If the linear vector field S = ∑ni=1 λi xi ∂∂xi is diophantine, then any holomorphic non-linear perturbation of S is holomorphically linearizable. This arithmetical condition has been weakened by A.D. Brjuno as we shall see below.

1.4 Examples about nonlinearizable vector fields Let’s go back to Example 1.2.3 where we saw that any holomorphic perturbation of order ≥ 3 of the normal form is formally conjugate to it. What about the holomorphy of such a conjugacy? Theorem 1.4.1 (Poincaré–Picard) If the linear part S has non-polynomial first integral but the constants and if there are no small divisors then any nonlinear perturbation X = S + R is holomorphically conjugate to a polynomial normal form in a neighborhood of the origin. Remark 1.4.2 Usually in the literature, the previous theorem is applied for linear part which spectrum is said to lie in the “Poincaré domain”. By this, we mean that there exists a line (D) in the complex plane which separate the eigenvalues of S from the origin (i.e. the eigenvalues are on one and the same side of the line while 0 is in the other side). Thus, if S belongs to the Poincaré domain, then it has only constant polynomial first integral. In fact, if LX (xQ ) = 0 then (Q, λ ) = 0. This means that the origin is a linear combination of the λi ’s with non negative coefficients. Since the spectrum lies on the same and opposite side from the origin of a line, then Q = 0. Furthermore, there are no small divisors since the projection of the eigenvalues onto the orthogonal line to (D) passing through the origin is bounded from below. So do the small divisors. Example 1.4.3 Let us show that Example 1.2.3 falls into the application scope of the theorem. In fact, a monomial (p,q are non-negative) x p yq is a first integral of S

Normal form of holomorphic dynamical systems

259

if and only if S(x p yq ) = (2p + q)x p yq = 0. This implies that p = q = 0. Thus, polynomial first integrals of S are just constants. Moreover, there are no small divisors. In fact, both |2p + q − 2| and |2p + q − 1| are integers so they don’t accumulate the origin. Therefore, any holomorphic system (6) is holomorphically conjugate to its normal form (7).

2 Holomorphic normalization The main progress are due to Brjuno who gave sufficient conditions that ensure that there is a convergent normalizing transformation to a normal form. These conditions are of two different type. The first one is a condition about the rate of accumulation to zero of the small divisors of the linear part. It is weaker that Siegel condition and is called condition (ω ). The second one is linked to the nonlinearity of the perturbation we are considering. It is a condition about a formal normal form of the perturbation.

2.1 Theorem of A.D. Brjuno Let X = S + R be an holomorphic vector field in a neighborhood of its singular point 0 ∈ Cn with S = ∑ni=1 λi xi ∂∂xi and R a nonlinear vector field. We assume that the following diophantine condition like is satisfied: (ω )

ln ωk < +∞ k k≥0 2

−∑

where ωk = inf{|(Q, λ ) − λi | = 0, 1 ≤ i ≤ n, Q ∈ Nn , 2 ≤ |Q| ≤ 2k }. Theorem 2.1.1 [Bru72] Let X = S +R be an holomorphic vector field as above. We assume that S satisfies the Bruno condition (ω ). If X has formal normal form of the type a.S ˆ for some formal power series aˆ (with a(0) ˆ = 1), then X is holomorphically normalizable. In the case of Hamiltonian vector field and under Siegel diophantine condition, this result is due to H. Rüssmann : Theorem 2.1.2 [Rüs67] Let H = ∑ni=1 λi xi yi + · · · be an analytic third order perturbation of the quadratic hamiltonian h = ∑ni=1 λi xi yi . Assume that h satisfies the Siegel condition:    n  c   µ  ∑ q jλ j >
 j=1  ∑nj=1 |q j |

260

L. Stolovitch

for integer vectors (q1 , . . . , qn ) ∈ Zn such that ∑nj=1 |q j | > 0. Assume that H has a ˆ ˆ ∑ni=1 λi xi yi ) then H is analytiformal Birkhoff normal form of the form F(h) = F( cally conjugate to a Birkhoff normal form F(h) for some analytic function F. We refer to J. Martinet’s Bourbaki seminar for a survey on this topic [Mar80]. Example 2.1.3 Let us apply the previous result to example (8). If it has a formal normal form (9) with Gˆ = Fˆ then it is holomorphically normalizable. Example 2.1.4 Let us consider the two-dimensional system $ x˙ = x2 y˙ = x + y

(1)

There is a unique formal diffeomorphism x = X, y = Y + ψˆ (X) that transforms the previous systems into its normal form $ x˙ = x2 (2) y˙ = y In fact, the conjugacy equation leads to y˙ = x + y = X +Y + ψˆ (X) = Y˙ + ψˆ (X)X˙ = Y + ψˆ (X)X 2 . So ψˆ has to solve the Euler equation X 2 ψˆ (X) − ψˆ (X) = X which formal solution is

ψˆ (X) = − ∑ (k − 1)!X k . k≥1

This does not converge in a neighborhood of the origin ! The normal form (2) does not satisfies Brjuno condition: it is not proportional to the linear part x˙ = 0, y˙ = y. Nevertheless, we can show that there exists sectorial normalizations. This means that there exists germs of holomorphic diffeomorphisms defined only in the product of sector with an edge at the origin (in the x plane) and a disc around 0 (in y) which conjugate equation (1) into its normal form. This the starting point of a long story that have been developed by J. Martinet and J.-P. Ramis [MR82, MR83] for two-dimensional vector fields and by J. Ecalle, S. Voronin and B. Malgrange for germs of local diffeomorphisms near a fixed point in the complex plane [Eca, Vor81, Mal82, Il 93]. In higher dimension, the theory has been developed by J. Ecalle and L. Stolovitch [Eca92, Sto96]. Recently, the interplay beetwen these “Stokes phenomena” and small divisors phenomena have been investigated by B. Braaksma and L. Stolovitch [BS07]. We refer to [Bal00, Ram93, RS93] for summability theory and Stokes phenomenon.

Normal form of holomorphic dynamical systems

261

2.2 Theorems of J. Vey On the other hand, Vey proved two theorems about the normalization of family of commuting vector fields satisfying some geometric properties. Theorem 2.2.1 [Vey79] Let X1 , . . . , Xn−1 be n − 1 holomorphic vector fields in a neighborhood of 0 ∈ Cn , vanishing at this point. We assume that : • Each Xi is a volume preserving vector field (LXi ω = 0 with ω an holomorphic n-differential form) • The 1-jet J 1 (X1 ), . . . , J 1 (Xn−1 ) are diagonal and independent over C (this means 1 that if there are complex constants ci such that ∑n−1 i=1 ci J (Xi ) = 0, then ci = 0 for all i.) • [Xi , X j ] = 0 for all indices i, j Then, X1 , . . . , Xn−1 are holomorphically and simultaneously normalizable. Theorem 2.2.2 [Vey78] Let X1 , . . . , Xn be n holomorphic vector fields in a neighborhood of 0 ∈ C2n , vanishing at this point. We assume that : • Each Xi is an Hamiltonian vector field • The 1-jet J 1 (X1 ), . . . , J 1 (Xn ) are diagonal and independent • [Xi , X j ] = 0 for all indices i, j Then, X1 , . . . , Xn are holomorphically and simultaneously normalizable.

2.3 Singular complete integrability–Main result We shall present a general result about normalization of commutative family of holomorphic vector fields vanishing at the same point that unifies both Vey’s and Brjuno’s theorems. At first glance, such unification could seem a little bit weird. In fact, in Vey’s theorems, there no assumption about small divisors while in Brjuno’s theorem there is one. In Vey’s theorem, vector fields satisfy a geometric assumptions (volume preserving or symplectic) whereas in Brjuno’s theorem there is an assumption about the formal normal form. Let us consider the family S = {S1 , . . . , Sl }, l ≤ n, of linearly independent linear diagonal vector fields n ∂ . Si = ∑ λi, j x j ∂ xj j=1 This means that if ∑li=1 ci Si = 0 for some complex numbers ci , then all the ci ’s are zero. Let us define the sequence of positive numbers $ % i n k ωk (S) = inf max |(Q, λ ) − λi, j | = 0, 1 ≤ j ≤ n, Q ∈ N , 2 ≤ |Q| ≤ 2 , , 1≤i≤l

where λ i = (λi,1 , . . . , λi,n ).

262

L. Stolovitch

Definition 2.3.1 We shall say that S is diophantine if (ω (S))

ln ωk (S) < +∞ 2k k≥0

−∑

Remark 2.3.2 The family S can be diophantine while none of the Si ’s satisfies Brjuno condition (ω ). For instance, consider in (C3 , 0) with complex coordinates (x, y, z) the vector fields S1 = E1 − ζ E2 and S2 = −ζ E1 + E2 where ζ is positive irrational number, E1 = x ∂∂x − y ∂∂y and E2 = y ∂∂y − z ∂∂z . Since 1 − ζ 2 = 0, S1 and S2 are linearly independent. The small divisors relative to S1 or S2 look like q1 − ζ q2 for some relative integers q1 , q2 . Thus, if ζ is a Liouvillian number then neither S1 nor S2 will satisfy Brjuno condition. On the other hand, let λi (resp. µi ) be the vector of eigenvalues of Si (resp. Ei ). We have      (Q, λ1 ) − λ1, j (Q, µ1 ) − µ1, j 1 −ζ A := = =: B. −ζ 1 (Q, λ2 ) − λ2, j (Q, µ2 ) − µ2, j Hence, if we denote the matrix by C, we have then B ≤ C−1 A. Therefore, the sequence of the A’s when Q and j vary do not accumulate the origin since the sequence of the B’s does not. So, the family S is diophantine. S 
S 
.1 resp. be the formal centralizer of S (resp. the ring of formal O-n Let X n first integrals), that is the set of formal vector fields X (resp. formal power series f ) such that [Si , X] = 0 (resp. LSi ( f ) = 0) for all 1 ≤ i ≤ l. Let X = {X1 , . . . , Xl } be a family of germs of commuting vector fields at the origin such that the linear part of Xi is Si ; that is [Xi , X j ] = 0 for all i, j. We shall call X a nonlinear deformation of S. Definition 2.3.3 We shall say that a nonlinear deformation X of S is a normal form (with respect to S) if [Si , X j ] = 0, 1 ≤ i, j ≤ l. Definition 2.3.4 We shall say that X, a nonlinear deformation of S, is formally completely integrable if there exists a formal diffeomorphism Φˆ fixing the origin and tangent to the identity at that point which conjugate the family X to normal form of the type

Φˆ ∗ Xi =

l

∑ aˆi, j S j ,

i = 1, . . . , l

(3)

j=1

where the aˆi, j ’s belongs to O-nS . Proposition 2.3.5 If X has a formally completely integrable normal form then all its normal form are also formally completely integrable. Theorem 2.3.6 Under the assumptions above, if S is diophantine, then any formally completely integrable nonlinear deformation X = S + ε of S is holomorphically normalizable.

Normal form of holomorphic dynamical systems

263

This means that there exists a genuine germ of biholomorphism Φ : (Cn , 0) → tangent to the identity at 0 which conjugate the family X to normal form of the type

(Cn , 0)

Φ∗ Xi =

l

∑ ai, j S j ,

i = 1, . . . , l

(4)

j=1

where the ai, j ’s are germ of holomorphic invariant functions, i.e. they belong to OnS . Remark 2.3.7 The theorem doesn’t says that neither Φˆ nor the aˆi, j converge but rather that there is another normalizing diffeomorphism that converges. Remark 2.3.8 One way to use this theorem is to have “a magic word in hand” (like Hamiltonian, volume preserving, reversible ....) that will implies that the formal normal form is of the good type. This comes from the data of the problem that one wants to solve. Corollary 2.3.9 If S is diophantine and if the holomorphic nonlinear deformation X is formally linearizable then it is holomorphically linearizable, i.e. there exists a holomorphic change of coordinates in which all the Xi ’s are linear. Of course, if one of the Si ’s satisfies Brjuno condition (ω ) and if the family X is formally linearizable, then it is also holomorphically linearizable. The point of the previous corollary is that none of the Si ’s is required to satisfies (ω ) in order that S to be diophantine. A result similar to our corollary was obtained by T. Gramchev and M. Yoshino for germs of commuting diffeomorphisms near a fixed point [GY99] under a slightly coarser diophantine condition. The article of J. Moser [Mos90] was the starting point since he was dealing with germ of one-dimensional diffeomorphisms.

2.3.1 Fundamental structures Proposition 2.3.10 [Sto00][prop. 5.3.2] With the notation above, O-nS is a formal .S is a O-S -module of finite type. C-algebra of finite type; X n n This means the following : if the ring of invariants is nor reduced to the constants, then there exists a finite number of monomials xR1 , . . . , xR p such that O-nS = C[[xR1 , . . . , xR p ]]. Moreover, there exists a finite number of polynomial vector fields .S (i.e. [Si , X] = 0, for all i) then there exY1 , . . . ,Ym such that if X belongs to X n S ists aˆ1 , . . . aˆm ∈ On such that X = aˆ1Y1 + · · · + aˆmYm . The proof is based on Hilbert theorem : in a Noetherian ring, ideals are generated by a finite number of elements. Let 2 ≤ k be an integer and let Pnk be the space of homogeneous vector fields of n C of degree k. Let us consider the map ρ : Cl → HomC (Pnk , Pnk ) defined by  

ρ (g)(X) =

l

∑ gi Si , X

i=1

264

L. Stolovitch

π −1 (0)

π −1 (c) π −1 (c )

0

NF2 NF1

π

Cπ ⊂ C p Fig. 1 Singular complete integrability: in the new holomorphic coordinate system, all the fibers (intersected with a fixed polydisc) are left invariant by the vector fields and their motion on it is a linear one

where g = (g1 , . . . , gl ) and X ∈ Pnk ([., .] denotes the Lie bracket of vector fields of Cn ). It is a representation of the commutative Lie algebra Cl in Pnk . To such a representation ρ of the abelian Lie algebra Cl into a finite dimensional vector space M, one can associate the Chevalley–Koszul complex

Normal form of holomorphic dynamical systems







265


 dl−1 d d d 0 → M →0 HomC Cl , M →1 HomC ∧2 Cl , M →2 · · · → HomC ∧l Cl , M → 0, (5)  where the differentials di are defined in the following way : if ω ∈ HomC ∧ p Cl , M and (g1 , . . . , g p+1 ) ∈ (Cl ) p+1 , then 



p+1

d p (ω )(g1 , . . . , g p+1 ) =

∑ (−1)i+1 ρ (gi ) (ω (g1 , . . . , g-i , . . . , g p+1 ))

(6)

i=1

Here (g1 , . . . , gˆi , . . . , g p+1 ) ∈ (Cl ) p stands for (g1 , . . . , gi−1 , gi+1 , . . . , g p+1 ). The differentials d0 and d1 will be particuliary useful: d0U(g) = ρ (g)U,

d1 F(g1 , g2 ) = ρ (g1 )F(g2 ) − ρ (g2 )F(g1 ).

The cohomology spaces H i (Cl , M) are defined to be H i (Cl , M) = Ker di /Im di−1 ,

i = 0, . . . l − 1.

Let α = (α1 , . . . , αl ) ∈ Cl . It defines the complex linear form on Cl , α (z) = To such a linear form, we associate the “generalized eigenspace”

Pn,k α = X ∈ Pnk | ∀g ∈ Cl , [S(g), X] = α (g)X .

∑li=1 αi zi .

In other words, X ∈ Pn,k α if and only if [Si , X] = αi X for all 1 ≤ i ≤ l. If Pn,k α = 0 then α is called a weight of S and Pn,k α is called the associated weightspace. There is a decomposition of the space into “generalized eigenspaces”, namely the Fitting decomposition: k k ⊕ Pn,0 Pnk = Pn,∗ k is the (finite) direct sum of the weightspaces associated to nonzero where Pn,∗ weights of S.

2.3.2 Geometric interpretation In order to illustrate our result, let us first recall the Liouville theorem [Arn97]. Let H1 , . . . , Hn be smooth functions on a smooth symplectic manifold M 2n ; let π : M 2n → Rn denotes the map π (x) = (H1 (x), . . . , Hn (x)). We assume that, for all 1 ≤ i, j ≤ n, the Poisson bracket {Hi , H j } = 0 vanishes. Let c ∈ Rn be a regular value of π ; we assume that π −1 (c) is compact and connected. Then there exists a neighborhood U of π −1 (c) and a symplectomorphism Φ from U to π (U) × Tn such that, in this new coordinate system, each symplectic vector field XHi associated to Hi is tangent to the fiber {d} × Tn . It is constant on it and the constant depends only on the fiber. Let us turn back to our problem and let S be a diophantine family of linearly independent diagonal vector fields of Cn . Let O-nS be its ring of formal first integrals. It is a C-algebra of finite type and there are homogeneous polynomials u1 , . . . , u p such

266

L. Stolovitch

that O-nS = C[[u1 , . . . , u p ]]. Let π : Cn → C p defined by π (x) = (u1 (x), . . . , u p (x)). Let s be the degree of transcendence of the field of fractions of C[u1 , . . . , u p ]; it is the maximal number of algebraically independent polynomials among u1 , . . . , u p . The algebraic relations among u1 , . . . , u p define an s-dimensional algebraic variety CS in C p . Hence, π defines a singular fibration over CS . The linear vector fields S1 , . . . , Sl are tangent and independent on each fiber π −1 (b) of π ; the latter are called toric variety because they admit an action of the algebraic torus C∗ . Note that we must have l ≤ n − s. Now, we come to the nonlinear deformation. Let X = S + ε be a nonlinear deformation of S. Let us assume that it is formally completely integrable. Then, according to our result, there exists a neighborhood U of 0 in Cn and an holomorphic diffeomorphism Φ on U such that, in the new coordinate system, the vector fields Φ∗ X1 , . . . , Φ∗ Xl are commuting linear diagonal vector fields on each fiber restricted to U and their eigenvalues depend only on the fiber. Indeed, in this new coordinates, we have Φ∗ Xi = ∑lj=1 ai, j S j where ai, j ∈ OnS . By definition, these vector fields are all tangent to the fibers of π (therefore, we must have l ≤ n − s). As consequence Φ∗ Xi ’s are all tangent to the fibers of π . On each fiber, the functions ai, j are constant so that each Φ∗ Xi reads as a linear diagonal vector field, that is a linear motion of a toric variety.

2.3.3 Proper Poincaré extension The next question that can be asked is the following : under what assumptions can a formally completely integrable nonlinear deformation X = S + ε of S be extended in an higher dimensional space into another formally completely integrable nonlinear ˆ with the same number of commuting vector fields? deformation Sˆ + εˆ of S, First of all, we shall define a good extension of S in Cn+m to be Sˆi := Si ⊕ Si

, i = 1, . . . , l, where Si

is a diagonal linear vector field of Pm1 . Of course, we want the properties of Sˆ to be derived from those of S; that is, we want Sˆ to be diophantine as Sˆ = OnS . One way to achieve this is to assume that soon as S is and we want that On+m

S is Poincaré family relatively to S : we require that the weights of S all belong to a real linear hyperplane of R2l whereas the weights of S

all, but a finite number of them, belong to one and the same side of the hyperplane. Such an extension will be called proper if the only weight of S

which belong to the hyperplane is the zero weight. If (x1 , . . . xn+m ) denotes coordinates of Cn+m and if X is a vector field of ∂ . (Cn+m , 0), then X

denotes its projection onto ∂ x∂ , . . . , ∂ xn+m n+1

Definition 2.3.11 We shall say that a proper Poincaré extension of S in Cn+m is completely integrable if there exists a formal diffeomorphism Φˆ fixing the origin and tangent to the identity at that point which conjugate the family X to normal form of the type NFi := Φˆ ∗ Xi =

l

l

j=1

j=1

∑ aˆi, j S j + ∑ aˆi, j S

j + D

i + Nili

+ Res

i ,

i = 1, . . . , l

(7)

Normal form of holomorphic dynamical systems

267

where the aˆi, j ∈ O-nS . Here, D

i (resp. Nili

, Res

i ) denotes a linear diagonal (resp. nilpotent, nonlinear) vector field of Cm with coefficient in O-nS such that the family D

has the same centralizer as S

(resp. commuting with the Si

’s). In other words, the projection NF

of the normal form is a polynomial Poincaré normal form of Cm with coefficients in OnS . Then we have the Theorem 2.3.12 [Sto00] Let S be a diophantine family of diagonal linear vector field of Cn . We assume that Sˆ = S ⊕ S

is a proper Poincaré extension of S in Cn+m by S

. Then, any nonlinear deformation of Sˆ which is formally completely integrable is holomorphically normalizable. For one vector field, theses results are due to Brjuno. Let us illustrate this result on Example 1.2.6. Let us define S = x1 ∂∂x − x2 ∂∂x − ζ x3 ∂∂x . Assume that S satisfies 1

2

3

Brjuno condition (ω ). Let us define S

= ix4 ∂∂x + ix5 ∂∂x . It is proper Poincaré vec4 5 tor field with respect to S. In fact, all the weights of S are real while those of S

are purely imaginary. Then nonlinear centralizer of S

is reduced to zero. First of all assume that in the normal form (10), we have fˆ1 = fˆ2 = fˆ3 . So that the projection on ∂∂x , . . . , ∂∂xn is a formally completely integrable system which nonlinearities 1 are parametrized by Cm . Assume that the formal power series gˆ1,1 and gˆ2,2 can be decomposed as gˆi,i = fˆi,i + hˆ i,i such that 1. q1 (i + fˆ1,1 (x1 x2 )) + q2 (i + fˆ2,2 (x1 x2 )) ≡ i + fˆ1,1 (x1 x2 ) q1 (i + fˆ1,1 (x1 x2 )) + q2 (i + fˆ2,2 (x1 x2 )) ≡ i + fˆ2,2 (x1 x2 ) for all (q1 , q2 ) ∈ N2 such that q1 + q2 ≥ 2. This precisely means that the formal vector field (i + fˆ1,1 (x1 x2 )x4 ∂∂x + (i + fˆ2,2 (x1 x2 )x5 ∂∂x , thought as a vector field 4 5 of C2 , has the same nonlinear centralizer as S

, that is 0. 2. The vector field 

hˆ 1,1 (x1 x2 )x4 + gˆ1,2 (x1 x2 )x5

 ∂  ∂  + gˆ2,1 (x1 x2 )x4 + hˆ 2,2 (x1 x2 )x5 ∂ x4 ∂ x5

is nilpotent and commutes with S

. Let us a give a geometric interpretation of this last result. Let us consider again the map π˜ : Cn+m → C p with π (x) = (xR1 , . . . , xR p ). Since, the invariants of Sˆ are the same as those of S, we have π˜ −1 (b) = π −1 (b) × Cm . Let us apply our result. In a new holomorphic coordinate system at the origin, the projection Xi on Cn of the vector field Xi is a completely integrable in the previous sense : it is tangent to any the toric variety π −1 (b) and its restriction to it is a linear diagonal motion. On the other hand, the projection Xi

on Cm is a polynomial normal form (of Cm ) which coefficients depend only holomorphically on b.

268

L. Stolovitch

2.4 How to recover Brjuno’s and Vey’s theorems from Theorem 2.3.6 Brjuno’s theorem correspond precisely to our result for l = 1. Let us prove the volume preserving case of Vey’s theorem. Let E be the family of the n − 1 linear semi-simple vector fields of Cn defined to be Ei = xi ∂∂xi − xi+1 ∂ x∂ , i+1 1 ≤ i ≤ n−1. The weights associated to Q = (q1 , . . . , qn ) ∈ Nn , |Q| ≥ 2, 1 ≤ j ≤ n are αi,Q, j = qi −qi+1 + δi, j δi+1, j (−1)δi, j (the last expression in the sum is 0 if j = i, i+1, 1 if j = i + 1 and −1 if j = i). First of all, the values of the nonzero weights of E are integers; thus, they cannot accumulate the origin, so that E is diophantine. E
.1 is the C[[u]]-module Moreover, if we set u = x1 · · · xn , then O-nE = C[[u]] and X n E
∂ .1 satisfies generated by xi ∂ xi , 1 ≤ i ≤ n. An easy computation shows that X ∈ X n to LX (u) = 0 if and only if X belongs to the C[[u]]-module generated by the Ei ’s. Let us write J 1 (Xi ) = ∑nj=1 µi, j xi ∂∂x j . Let us set µ i := (µi,1 , . . . , µi,n ). Since Xi is volume preserving then, µi,1 + · · ·+ µi,n = 0; it follows that J 1 (Xi ) = ∑n−1 j=1 ai, j E j . By the independence of the 1-jets, the (n − 1) × (n − 1) matrix A0 = (ai, j ) is invertible. Let us compute the weights of the family of the J 1 (Xi )’s with respect of those of E. We have ⎞ ⎛ ⎞ ⎛ (Q, µ 1 ) − µ1, j α1,Q, j ⎟ ⎜ . ⎟ ⎜ .. ⎠ = A0 ⎝ .. ⎠ . ⎝ .

αl,Q, j

(Q, µ l ) − µl, j Therefore, we have

     (Q, µ 1 ) − µ1, j   α1,Q, j         ..  ..  ≥ A−1   .  0 .    .   (Q, µ l ) − µl, j   αl,Q, j  This means that the family of the J 1 (Xi )’s is also diophantine. Since the family can be transformed into a normal form, there exists a formal diffeomorphism Φˆ such that Φˆ ∗ Xi = ∑nj=1 Fˆi, j (u)xi ∂∂xi for some Fˆi, j ∈ C[[u]]. We can assume that Φˆ is volume preserving; thus the normal forms are also volume preserving. Hence div (Φˆ ∗ Xi ) = 0, that is 
  n n d ∑nj=1 Fˆi, j ∂ x j Fˆi, j (u) (u). ∑ ∂ x j = 0 = ∑ Fˆi, j (u) + u du j=1 j=1 An easy computation shows that ∑nj=1 Fˆi, j = 0. Thus,

Φˆ ∗ Xi =

n−1

n−1

j=1

j=1

∑ fˆi, j (u)E j = ∑ gˆi, j (u)J 1 (X j ),

Normal form of holomorphic dynamical systems

269

that is X is formally completely integrable. According to our main result, there is an holomorphic diffeomorphism Ψ normalizing X in a neighborhood of the origin. By a classical argument of Vey [Vey79], we can modify holomorphically Ψ so that it becomes volume preserving and still normalizing X.

2.5 Sketch of the proof Let us give a sketch of the proof of our results. In order to normalize the nonlinear deformation X = S + ε of S, we shall proceed through a classical Newton method, that is a Nash–Moser induction type. Let us assume that the nonlinear deformation X = S + ε is normalized up to order m; we will build a diffeomorphism Φm which normalize the deformation up to order 2m; it is tangent to Id up to order k. Let us show how this works. First of all, we can write the deformation Xi = NFim + Bi + Ri , 1 ≤ i ≤ l where NFim is a normal form of degree m, Bi is polynomial of degree ≤ 2m and of order ≥ m + 1 and Ri is of order ≥ 2m + 1. Let us denote by Bi,∗ (resp. Bi,0 ) the projection of Bi onto the sum of the weightspaces associated to a nonzero weight (resp. zero weight) of S in Pnm+1,2m . The compatibility condition (i.e [Xi , X j ] = 0 for all 1 ≤ i, j ≤ l) shows that, for all 1 ≤ i, j ≤ l   (8) J 2m [NFim , B j,∗ )] − [NFjm , Bi,∗ ] = 0. On the other hand, if we conjugate Xi by a diffeomorphism of the form exp(U) for some polynomial vector field U ∈ Pnm+1,2m and writing exp(U)∗ Xi = NFim +B i +R i as above, we find out that   J 2m B i − Bi + [NFim ,U] = 0 The algebraic properties of the weightspaces of S show that, in fact, we have   J 2m B i,∗ − Bi,∗ + [NFim ,U∗ ] = 0. If we assume that the diffeomorphism exp(U) normalizes simultaneously the Xi ’s up to order 2m then we must have B i,∗ = 0 for all i. Hence, we have J 2m (−Bi,∗ + [NFim ,U∗ ]) = 0 i = 1, . . . , l.

(9)

m+1,2m the direct sum of weightspaces associated to a nonzero Let us denote by Pn,∗ weight of ρ in Pnm+1,2m . Let us define the linear map  m+1,2m  m+1,2m ρm : Cl → HomC Pn,∗ , Pn,∗

270

by ρm (g)(X) = J 2m

L. Stolovitch


!

∑li=1 g j NFjm , X

" if g = (g1 , . . . gl ). It is well defined and it

m+1,2m . To this representais a representation of the abelian Lie algebra Cl into Pn,∗ tion is associated a complex of finite dimensional complex vector spaces; it is the Chevalley-Kozsul complex of this representation. Let us write dmi its ith-differential. Therefore, equation (8) reads dm1 (B∗ ) = 0, that is B∗ is a 1-cocycle for this complex. Moreover, equation (9) reads dm0 (U) = B∗ , that is B∗ is the 0-coboundary of U: it is a cohomological equation. Hence, the Chevalley–Koszul complex of the representation ρm plays an important role in our problem. We shall call it the Newton complex of order m. According to the discussion above, the first important problem to study is its cohomology. We can show that its 0th-cohomology as well as the 1st-cohomology spaces are zero:

Proposition 2.5.1 [Sto00][Prop. 7.1.1] We have
 m+1,2m = 0, i = 0, 1 Hmi Cl , Pn,∗ where Hmi denotes the ith-cohomology space of the Chevalley–Koszul complex associated to ρm . It is not very difficult but rather technical. It leads to the important consequence m+1,2m such that, for all that, B∗ being given as above, there exists a unique U ∈ Pn,∗ 2m m 1 ≤ i ≤ l, J ([NFi ,U]) = Bi,∗ ; hence, conjugating Xi by exp(U) normalizes Xi up to order 2m. We find out that the formal diffeomorphism defined by Φˆ := limk→+∞ Φ2k ◦ · · · ◦ Φ2 normalizes simultaneously the Xi ’s where the Φ2k ’s are built as above. In order to prove that Φˆ is holomorphic in a neighborhood of 0 ∈ Cn , one has to estimate Φ2k . Here comes the analysis and the major difficulty. To get an estimate of Φm = exp(U) with m = 2k , we have to estimate U. Hence, we are led naturally to give bounds for the cohomology of the Newton complex : Let r > 1/2, the spaces of the Newton complex are provided with norms (depending on a real positive number r) which turn it into a topological complex of vector spaces. By the above algebraic properties, the 0-differential, dm0 , has a right inverse s on the space of 1-cocycle : if Z is m+1,2m a 1-cocycle of the Newton complex, then s(Z) is the unique element of Pn,∗ 0 such that dm (s(Z)) = Z. Here comes the main assumptions : if the family X is completely integrable then there exists constants d, η1 , c(η1 ), such that if m = 2k and if the r-norms of NF m − S and D(NF m − S) are sufficiently small, say < η1 (for some 1/2 ≤ r < 1) then |s(Z)|r ≤

c(η1 ) |Z|r ; d (S) ωk+1

(10)

the constant d doesn’t depend on η1 (we recall that ωk (S) is the smallest norms of k

the nonzero weights of S in Pn2,2 ). Let us describe the way we obtain this estimate. In order to solve the cohomological equation associated the 1-cocycle Z, it is necessary and sufficient to solve the

Normal form of holomorphic dynamical systems

271

system of l equations J 2m ([NFim ,U]) = Zi , i = 1, . . . , l. We can decompose this equation along the weightspaces of S. In fact, let α be a weight of S and let V belongs to the associated weightspace. Then, by Jacobi identity, we have [S j , [NFim ,V ]] = [−NFim , [V, S j ]] − [V, [S j , NFim ]]. m−1 By assumptions, NFim = ∑lj=1 am−1 i, j S j where the ai, j ’s are polynomials invariants for S of degree ≤ m − 1. Therefore, according to formula (2), [S j , NFim ] = 0. Hence, we have [S j , [NFim ,V ]] = [−NFim , [V, S j ]] = α j [NFim ,V ]. It is sufficient to consider for any nonzero weight α of S, the equation with both Zi ’s and U in the associated weightspace. This set of equations can be written in the following matrix form ⎞ ⎞ ⎛ ⎞ ⎛ ⎛ α1U D1 (U) Z1 + Z1 ⎟ ⎟ ⎜ ⎟ ⎜ ⎜ A(x) ⎝ ... ⎠ + ⎝ ... ⎠ = ⎝ ... ⎠

αl U

Dl (U)

Zl + Zl

S where A = (am−1 i, j ) is a square l × l matrix with coefficients in the C-algebra On of holomorphic first integrals of the linear part S; A(0) = Id; the operators D1 , . . . , Dl are OnS -linear; Z1 , . . . , Zl have order ≥ 2m + 1. After inverting the matrix A, we obtain l equations (αi Id + D˜ i )(U) = Z˜ i + Z˜ i , i = 1, . . . , l. The D˜ i ’s (resp. Z˜ i ,Z˜ i ) are still OnS -linear operators and they are linear combination of the Di ’s (resp. Zi , Zi ) with coefficients in OnS . Let us set α  = max1≤ j≤l |α j | and let i be such that |αi | = α  = 0; it is the “worst small divisor” of the family. Let us look through the ith equation; we find out that, at least formally, its solution U is given by   −1 k ˜ k ˜ ˜ 1 U= ∑ αi Di (Zi + Zi ). αi k≥0

This expression does not fancy us since it involves a priori infinitely large powers of αi which can be very small. Thus, instead of using this expression, we shall split the ith equation in an appropriate way. First of all, we shall split the linear diagonal

family S in two parts S and S

corresponding to the splitting of Cn as Cn × Cn−n ; that is, for all 1 ≤ i ≤ l, n

Si =

∂

n

∑ λi,k xk ∂ xk + ∑

k=1

3

45 Si

λi,k xk

+1 6 k=n 3 45 Si

∂ . ∂ xk 6

The integer n is such that the linear forms {∑li=1 λi,k zi }1≤k≤n all belong to a real hyperplane H of HomC (Cl , C) whereas all the linear forms {∑li=1 λi,k zi }n +1≤k≤n all belong (strictly) to one and the same side of H. The integer n is taken to be the lowest as possible; it may be equal to 0 as well as equal to n. We shall call this splitting, the analytic splitting of S. It has been chosen in such a way that the

272

L. Stolovitch

small divisors as well as the first integrals are only due to S . We show that there is a separating constant Sep(S) > 0 such that if α is a weight of S which norm is < Sep(S) then it must belong to H (if n = n we shall set Sep(S) = +∞ in order to have one proof for the theorems). Let X be a vector field of Cn , we shall denote by X (resp. X

) its projection onto ∂∂x , . . . , ∂ ∂x (resp. ∂ x ∂ , . . . , ∂∂xn ). This being 1 n n +1 said, let us go back to the study of our equation (αi Id + D˜ i )(U) = Z˜ i + Z˜ i . Using the analytic splitting of S as well as the structure of the operator D˜ i , we show that this equation can be written under the following form : 1 1 (Pi (U )) = (Z˜ i + Z˜ i + (Qi (U )) ) αi αi 1 1





U − (Qi (U )) = (Z˜ i

+ Z˜

i + (Pi (U ))

+ (Qi (U ))

); αi αi U −

(11) (12)

both Pi and Qi are OnS -linear operators. Let us assume that the weight α is of small norm, that is < Sep(S). Then, we show that (Qi (U )) = 0 and that, according to the complete integrability assumption, Pi ◦ Pi = 0. Therefore the solution of (11) is given by  ˜   Z˜ i + Zi Pi

. U = Id + αi αi Since U is a polynomial of order ≤ 2m, then in fact, we have      Pi Z˜ i 

 |U |r ≤  Id + . αi αi r An estimate of the operator Pi will provide the desired estimate of U . Now, let us study equation (12). Let us denote by α1i wi the left handside of this equation. Then, at least formally, we have U

=

∑

k≥0



1 αi

k

 Qki

wi αi

 .

By assumption, NF m is the m-jet of completely integrable normal form. Therefore, its projection (NF m )

is the m-jet of a good deformation of S

. The point is that there exists an integer k0 which do not depend on m and such that i J 2m (Qki ( w αi )) = 0 for all k ≥ k0 . The important consequence for the estimates is that the sum above which give U

is finite. Using the estimate of U which were found above, we can give estimate for wi ; then using estimate of Qi , we conclude with estimate of U

. The last case deals with weight α such that α  ≥ Sep(S); it it the easiest case. Now let us give an idea of the induction argument. Let 1/2 ≤ r < 1 and let assume the family of the Xi = NFim + Ri,m+1 ’s is normalized up to order m = 2k . Let us assume that the norms |NF m − S|r and |D(NF m − S)|r are small enough, say < η1 , and that |Ri |r < 1. The solution of the cohomological equation allows us to normal-

Normal form of holomorphic dynamical systems

273

ize the family up to order 2m : (Φm )∗ Xi = NFi2m +Ri,2m+1 . Using the estimate of this solution, we show that |NF 2m − S|R and |D(NF 2m − S)|R are still less than η1 where  −1/m η1 ) R = c( m−2/m r < r and that |Ri,2m+1 |R < 1. After a preliminary renorωd k+1

malization, we show that, at each stage, our new objects still satisfy the required assumptions in order to have again the estimate for the solution of the new cohomological equation. Thus, we may proceed again ... Now, because of the diophantine condition, these R are bounded from below by some positive constant Rad. Therefore, at the limit, we have found an holomorphic diffeomorphism in the polydisc of radius Rad centered at 0 ∈ Cn which normalizes our nonlinear deformation X.

3 Proof of main Theorem 2.3.6 3.1 Bounds for the cohomological equations Let α be a nonzero weight of S in Pnm+1,2m and let Pn,m+1,2m be the associated α 1 weight space. As we have seen in proposition 2.5.1, for all Z ∈ ZN,m (Cl , Pn,m+1,2m ), α there exists a unique U ∈ Pn,m+1,2m such that, for all integer 1 ≤ i ≤ l, α J 2m ([NFim ,U]) = Zi . The remaining of this subsection is devoted to the determination of a bound of the norm of this solution under some assumptions. Moreover, we assume that NF m is the m-jet of the normal form of a completely integrable deformation of S. More precisely, we shall prove the Theorem 3.1.1 Under the assumptions above, there exists constants η1 > 0 and c1 (η1 ) > 0 such that, if 1/2 < r ≤ 1, m = 2k and max(|NF m −S|r , |D(NF m − S)|r ) < η1 , then for any nonzero weight α of S in Pnm+1,2m , for any Z ∈ Zm1 (Cl , Pn,m+1,2m ), α 0 U = Z satisfies the following inequality: the unique U ∈ Pn,m+1,2m such that dN,m α |U|r ≤

c1 (η1 ) |Z|r ; ωk+1 (S)2

(1)

and d depends only on S. Proof. The cohomological equation can be written: [NFim ,U] = Zi + Zi ,

i = 1, . . . , l.

where we have set, for all integers 1 ≤ i ≤ l, Zi := [NFim ,U] − J 2m ([NFim ,U]). By S assumptions, we have for all 1 ≤ i ≤ l, NFim = ∑lj=1 am−1 i, j S j where ai, j ∈ On are polynomials of degree ≤ m − 1 and am−1 i, j (0) = δi, j . Therefore, we have

274

L. Stolovitch

[NFim ,U] =

l

∑




 m−1 am−1 = Zi + Zi , [S ,U] −U(a )S j j i, j i, j

i = 1, . . . l

(2)

j=1 m−1 where U(am−1 i, j ) denotes the Lie derivative of ai, j along U. Let us choose an index 1 ≤ i ≤ l such that |α (gi )| = α  = 0. We recall that U belongs to the α -weightspace of S in Pnm+1,2m ; that is, for all 1 ≤ i ≤ l, [Si ,U] = αiU. Therefore, equations (2) can be written into the following matricial form ⎞ ⎞ ⎛ ⎞ ⎛ ⎛ [S1 ,U] D1 (U) Z1 + Z1 ⎟ ⎟ ⎜ ⎟ ⎜ ⎜ A(x) ⎝ ... ⎠ + ⎝ ... ⎠ = ⎝ ... ⎠

[Sl ,U]

Dl (U)

Zl + Zl

where A = (am−1 p,q )1≤p,q≤l and Di is the On -linear map defined by Di : U ∈ m−1 .2 → − ∑l U(a .2 X n j=1 i, j )S j ∈ X n . Since A(0) = Id, A(x) is formally invertt ˜ ible : if A := (ci, j )1≤i, j≤l denotes the transpose of the cofactors matrix of A, then Aˆ −1 := det1 A A˜ t := (bi, j )1≤i, j≤l is a matrix which coefficient belong to O-nS and satisfy to Aˆ −1 A = AAˆ −1 = Id. It follows that ⎛ ⎞ ⎞ ⎛ ⎞ α1U Z1 + Z1 D˜ 1 (U) l ⎜ .. ⎟ ⎜ .. ⎟ ˆ −1 ⎜ .. ⎟ ⎝ . ⎠ + ⎝ . ⎠ = A ⎝ . ⎠ where D˜ j (U) = ∑ b j,k Dk (U). k=1 Zl + Zl αl ,U D˜ l (U) ⎛

Here is a key point : equation (2) is overdetermined. To estimate its solution, we select the equation that give the smallest norm a priori. It is the one that correspond to the “biggest” small divisor among the family, that is αi . Thus, the ith equation of the cohomological equation can be written U − Pi (U) = Z˜ i + Z˜ i

(3)

Here, we have written

αi Z˜ i =

l

∑ bi,k Zk ,

αi Z˜ i =

k=1

l

Pi (U) =

l

∑ bi,k Zk ,

k=1 l

1 ∑ bi,k ∑ U(am−1 k, j )S j . αi k=1 j=1

We claim that the operator Pi satisfies to Pi ◦ Pi = 0. This is due to the fact that m−1 Sq (am−1 k,p ) = 0 since the ak,p ’s are invariants of S. Hence, we have (Id − Pi ) ◦ (Id + Pi ) = Id. As a consequence, we have U = (Id + Pi )(Z˜ i + Z˜ i ).

(4)

Normal form of holomorphic dynamical systems

275

Let us give bounds for the operators Pi . To do so, we shall write A(x) = Id + R(x) where R(0) = 0; we shall write R(x) = (ri, j (x))1≤i, j≤l . Recalling the expression of the determinant, we have det(A(x)) = 1 + P(R(x)) where P(Z) ∈ C[Z1 , . . . , Zl 2 ] is a polynomial functions of l 2 variables without constant term and of degree l. Since, it vanishes at the origin, there exists η > 0 such that |P(Z)|η < 1/2. It follows that, if |R(x)|r = |A(x) − A(0)|r < η , then |P(R(x))|r < 1/2. By Lemma 1.1.2, if |A(x) − A(0)|r < η , then we have    1  1    det A(x)  ≤ 1 − |P(ri, j (x))|r ≤ 2 r 3 | det A(x)|r ≤ 1 + |P(R(x))|r ≤ . 2 We recall that (ci, j )1≤i, j≤l = A˜ t is the transpose of the cofactors matrix of A. Thus, qi, j,S Z S ∈ there are universal polynomials of degree ≤ l − 1, Qi, j (Z) = ∑ l2 S∈N 1≤|S|≤l−1

C[Z1 , . . . , Zl 2 ] such that ci, j (x) = Qi, j (A(x)). It follows that, for all 1 ≤ i, j ≤ l, |ci, j (x)|r ≤ Q(|A|r ) where Q is the universal polynomial of one variable defined by Q(t) = ∑ max |qi, j,S |t |S| . 2

S∈Nl 1≤|S|≤l−1

i, j

As a consequence, if |A(x) − A(0)|r < η , we have    c   1   i, j   |ci, j |r ≤ 2Q(|A|r ).  |bi, j |r =   ≤ det A r  det A r

(5)

In our result, the assumption are about |NFim − Si |r and |D(NFim − Si )|r and not on the matrix A. So we have to link both estimates. By definition, we have, for all integers 1 ≤ i, j ≤ l

∂

n

Sj =

∑ λ j,k xk ∂ xk

k=1

NFim

l

=

∑

ai,k−1 j Sj

j=1

that is

⎛

∂ := ∑ xk gi,k ∂ xk k=1



n

with

gi,k =

⎞ ⎛ ⎞ λ1,1 . . . λl,1 ⎛ k−1 ⎞ gi,1 ⎜ .. ⎟ ⎜ .. .. ⎟ ai,1 ⎜ . ⎟ ⎜ . . ⎟ ⎜ ⎟=⎜ ⎟ ⎜ .. ⎟ ⎜ . ⎟ ⎜ . .. ⎟ ⎝ . ⎠ . ⎝ .. ⎠ ⎝ .. . ⎠ ak−1 i,l gi,n λ1,n . . . λl,n

l

∑

j=1



λ j,k ak−1 i, j

,

276

L. Stolovitch

By assumptions, the Si ’s (1 ≤ i ≤ l) are linearly independents over C, the matrix (λ j,i )1≤i≤n has rank l. Without lost of generality, we may assume that the matrix 1≤ j≤l

L := (λ j,i )1≤i, j≤l is invertible with inverse L−1 := (λ˜ i, j )1≤i, j≤r . m−1 Let us set A = (am−1 i, j )1≤i, j≤l and |A|r = maxi, j |ai, j |r , |D(A)|r = maxi, j,k | Let We have the following

∂ am−1 i, j ∂ xk |r .

Lemma 3.1.2 |A − A(0)|r ≤ 2l|L−1 ||NF m − S|r |D(A)|r ≤ 2l|L−1 ||D(NF m − S)|r

(6) (7)

We refer to [Sto00][p. 185–186] for a proof. Let us set η1 = η /(2l|L−1 |). If |NF m − S|r < η1 , then by (6), we have |A(x) − A(0)|r < η so that |bi, j |r ≤ 2Q(|A|r ) by (5). Moreover, we have |A|r ≤ |A(0)| + |A − A(0)|r ≤ 1 + η . It follows that Q(|A|r ) < Q(1 + 2l|L−1 |η1 ). k−1 On the other hand, if |D(NF m − S)|r < η1 then |U(ak, j )|r ≤ n|U|r |D(A)|r ≤ nη |U|r . We recall (see the section of notations) that, given two elements Y = (Y1 , . . . ,Yq ) and W = (W1 , . . . ,Wq ) of O-nq , we say that Y is dominated by W , and we write Y ≺ W , if Yi ≺ Wi for all 1 ≤ i ≤ q. Moreover, we shall write Y¯ := (Y¯1 , . . . , Y¯q ). Now, we are able to give estimates for Pi . Since we have Pi (U) = we obtain Pi (U) ≺

l 1 l k−1 bi,k ∑ U(ak, ∑ j )S j , αi k=1 j=1

l 1 l ¯ ¯ a˜k−1 )S¯ j . bi,k ∑ U( ∑ k, j α  k=1 j=1

Here, S¯ j stands for ∑nk=1 |λ j,k |xk ∂∂x . It follows that if 1/2 < r ≤ 1, max(|NF m − k S|r , |D(NF m − S)|r ) < η1 , then |Pi (U)|r ≤ with

c2 (η1 ) |U|r α 

(8)

c2 (η1 ) = 4l 3 Q(|A(0)| + 2l|L−1 |η1 )nλ |L−1 |η1 .

Here we have set λ = max1≤i≤l,1≤ j≤n |λi, j |, so that, since r ≤ 1, |S j |r ≤ λ for all 1 ≤ j ≤ l. Let us give the estimate of the solution of the cohomological equation (4). Since U is a polynomial vector field of degree ≤ 2m then U ≺ (Id + P¯i )(Z˜ i ). Hence,     |U|r ≤ (Id + P¯i )(Z˜ i ) r

Normal form of holomorphic dynamical systems

277





c2 (η1 ) ˜ |Zi |r α    c2 (η1 ) 2lQ(1 + 2l|L−1 |η1 ) |Z|r . ≤ 1+ α  α  ≤

1+



we are done.

3.2 Iteration scheme Let 1/2 < r ≤ 1 be a real number and η1 > 0 be the positive number defined in Theorem 3.1.1. For any integer m ≥ [8n/η1 ] + 1, let us set % $ 8n N F m (r) = X ∈ (Pn1,m )l | max(|Xi − Si |r , |D(Xi − Si )|r ) < η1 − m

m+1 l Bm+1 (r) = X ∈ (Xn ) | |Xi |r < 1 . If m = 2k , for some integer k ≥ 1, let us define  −1/m

ρ =m

r

and

−2/m

R = γk m

r

where γk =

c1 2 (S) ωk+1

−1/m .

The core of this section is the following proposition : Proposition 3.2.1 With the notations above, let us assume that (NF m , Rm+1 ) ∈ N F m (r)×Bm+1 (r). If m is sufficiently large (say m > m0 indepenm+1,2m dent of r), then there exists a unique U ∈ Pn,∗ , the sum of nonzero weightspaces of S, such that 1. Φ := (Id +U)−1 ∈ Diff1 (Cn , 0) is a diffeomorphism such that DR ⊂ Φ (Dρ ) 2. Φ ∗ (φ + ε ) = NF 2m + R2m+1 is normalized up to order 2m 3. (NF 2m , R2m+1 ) ∈ N F 2m (R) × B2m+1 (R) m+1,2m + We have seen that Φ normalizes simultaneously each Xi = NFm i (y) + Bi 2m

Ci (y) up to order 2m. Let us write Φ∗ Xi (y) = NFi (y) + Ci (y) where both Ci and Ci is of order ≥ 2m + 1 whereas Bim+1,2m is of degree ≤ 2m and of order ≥ m + 1. Using the conjugacy equation and the fact that NFm (Φ −1 (y)) − NFm (y) = 1 m 0 D(NF )(y + tU(y))U(y)dt, we have

Ci (y) =

 1 0

m+1,2m D(NFm +Ci )(Φ −1 (y)) i )(y + tU(y))U(y)dt + (Bi

(9)

m 2m

−(NF2m i (y) − NFi (y)) − D(U)(y)(NFi +Ci )(y).

This is expression that we will use in order to prove the third point of the proposition.

278

L. Stolovitch

3.2.1 Estimate for the diffeomorphism By assumptions, NF m ∈ N F m (r); thus, we can apply Proposition 3.1.1 so that |U|r ≤

c1 |Bm+1,2m |r . 2 (S) ∗ ωk+1

m+1,2m m+1,2m m+1,2m ≺ B¯ i,∗ + B¯ i,0 ≺ R¯ i,m+1 , we have |B∗m+1,2m |r < 1. It follows Since Bi,∗ that |U|r ≤ γk−m .

Lemma 3.2.2 Under the above hypothesis and if m is large enough (say m > m0 ), then for all 0 < θ ≤ 1 and all integer 1 ≤ i ≤ n, we have |yi + θ Ui (y)|R < ρ . As a consequence, Φ (Dρ ) ⊃ DR . Proof. We borrow the proof of Bruno [Bru72][p. 203]. It is sufficient to show that R + |U|R < ρ . Since U is of order at least m + 1 then, by (4) and the inequality above |U|R ≤

 m+1 m+1
R |U|r ≤ γk m−2/m γk−m r

≤ γk m−2−2/m ≤ m−2−2/m .

(10)

Since R = γk m−2/m r ≤ m−2/m r, it is sufficient to show that m−2/m (r + m−2 ) < ρ = m−2 < r. But, m−1/m r, that is m1/m −1 1 m−2 m−2 m−2 ≤ = ≤ 1/m 1/m ln m m − 1 exp −1 1/m ln m m ln m since 1+x ≤ exp x for all x ∈ R+ . But, for 0 < x sufficiently large, we have 2 < x ln x. m−2 Thus, since 1/2 < r, we obtain the result : m1/m < 12 < r. −1

3.3 Proof of the theorem In this section, we shall prove our main result. Let 1/2 < r ≤ 1 be a positive number and let us consider the sequence {Rk }k≥0 of positive numbers defined by induction as follow: R0 = r Rk+1 = γk m−2/m Rk

where

m = 2k

Lemma 3.3.1 The sequence {Rk }k≥0 converges and there exists an integer m1 such Rm that for all integer k > m1 , Rk > 2 1 .

Normal form of holomorphic dynamical systems

 Proof. We recall that γk =

c1

−1/2k

2 (S) ωk+1

279

. We have Rk+1 = r ∏ki=1 γi (2i )−2

1−i

, thus

k k ln ωi+1 (S) 1 i − ln c − 2 ln 2 1 ∑ ∑ i i i 2 2 2 i=1 i=1 i=1 k

ln Rk+1 = −2 ∑

the last two sums are convergent and the first is also convergent by assumption. It follows that there exists an integer m1 , such that +∞

∏

γi (2i )−2

1−i

> 1/2.

i=m1 +1

Thus, if k > m1 then Rk = Rm1 ∏ki=m1 +1 γi (2i )−2

1−i

>

Rm1 2 .

Let X = {Xi = Si + · · · }i=1,...,l be a family of commuting holomorphic vector fields in a neighborhood of the origin in Cn . We may assume that it is holomorphic in a neighborhood of the closed polydisc D1 . Let m2 = 2k0 ≥ max(m0 , 2m1 ) where m0 is the integer defined in Proposition 3.2.1. By a polynomial change of coordinates, we can normalize X up to order m2 : in these coordinates, Xi can be written NFim2 + Ri ,m2 +1 . If necessary, we may apply a diffeomorphism aId with a ∈ C∗ sufficiently small so that (NF m2 , Rm2 +1 ) ∈ N F m2 (1) × Bm2 +1 (1). We may define as above the sequence {Rk }k≥k0 , with Rk0 = 1. Thus, for all integer k > k0 , we have 1/2 < Rk ≤ 1. Let us prove by induction on k ≥ k0 , that there exists a diffeomorphism Ψk of k+1 (Cn , 0) such that Ψk∗ (NFim2 + Ri,m2 +1 ) := NFi2 + Ri,2k+1 +1 is normalized up to order 2k+1 , (NF 2

k+1

, R2k+1 +1 ) ∈ N F 2k+1 (Rk+1 ) × B2k+1 +1 (Rk+1 ) and |Id − Ψk−1 |Rk+1 ≤

k

∑

p=k0

1 . 22p

• For k = k0 : According to Proposition 3.2.1, there exists a diffeomorphism Φk0 k k +1 such that Φk∗0 (NF 2 0 + R2k0 +1 ) = NF 2 0 + R2k0 +1 +1 is normalized up to ork +1

der 2k0 +1 , (NF 2 0 , R2k0 +1 +1 ) ∈ N F 2k0 +1 (Rk0 +1 ) × B2k0 +1 +1 (Rk0 +1 ). The es|Rk +1 = |U|Rk +1 < 1/22k0 comes from estimate (10) timate |Id − Φk−1 0 0 0 • Let us assume that the result holds for all integers i ≤ k − 1 : by assump∗ (NF m2 + R 2k k tions, Ψk−1 m2 +1 ) = NF + R2k +1 is normalized up to order 2 k

and (NF 2 , R2k +1 ) ∈ N F 2k (Rk ) × B2k +1 (Rk ). Since 1/2 < Rk ≤ 1, we may apply proposition 3.2.1 : there exists a diffeomorphism Φk such that (Φk ◦ k k+1 Ψk−1 )∗ (NF 2 0 + R2k0 +1 ) = NF 2 + R2k+1 +1 is normalized up to order 2k+1 k+1

and (NF 2 , R2k+1 +1 ) ∈ N F 2k+1 (Rk+1 ) × B2k+1 +1 (Rk+1 ). Let us set Ψk = Φk ◦ Ψk−1 . According to estimate (10), we have |Id − Φk−1 |Rk+1 < 1/22k . It follows that

280

L. Stolovitch

|Id − Ψk−1 |Rk+1

  −1 ≤ (Id − Ψk−1 ) ◦ Φk−1 + (Id − Φk−1 )R , k+1     −1 ≤ (Id − Ψk−1 ) ◦ Φk−1 R + (Id − Φk−1 )R

k+1

k+1

.

According to proposition (3.2.1), we have Φk−1 (DRk+1 ) ⊂ DRk . It follows that     −1  |Id − Ψk−1 |Rk+1 ≤ (Id − Ψk−1 ) R + (Id − Φk−1 )R k

≤

k−1

∑

p=k0

k+1

1 1 + ; 22p 22k

this ends the proof of the induction. Since D(1/2) ⊂ DRk for all integers k ≥ k0 , then the sequence {|Ψk−1 |1/2 }k≥k0 is uniformally bounded. Moreover, the sequence {Ψk−1 }k≥k0 converges coefficientwise to a formal diffeomorphism Ψˆ −1 (the inverse of the formal normalizing diffeomorphism). Therefore, this sequence converges in Hnn (r) (for all r < 1/2) to Ψˆ −1 (see [GR71]). This means that the normalizing transformation is holomorphic in a neighborhood of 0 ∈ Cn .

4 Miscellaneous results 4.1 Normal forms again The following theorem is due to Nguyen Tien Zung. It describes a situation of “fully” completely integrable systems : there is “no room” left for small divisors and there are the maximum number of holomorphic first integrals. Theorem 4.1.1 [Zun02] Let X1 = S + R be an holomorphic nonlinear perturbation of a semi-simple linear part S in a neighborhood of the origin in Cn . Assume it belongs to a commutative family a holomorphic vector fields {X1 , . . . , Xm } which are assumed to be linearly independent almost everywhere : X1 ∧ · · · ∧ Xl = 0 almost everywhere. Assume that this family has n − m holomorphic common first integrals { f1 , . . . , fn−m } which are almost everywhere independent : d f1 ∧ · · · ∧ d fn−m = 0 almost everywhere. Then, there exists an holomorphic change of coordinates which normalize X1 . For the Hamiltonian version of this result see [Zun05]. This result is very related to the following one. As above, let us consider the family S of linear vector fields Si = ∑nj=1 λi, j x j ∂∂x j , 1 ≤ i ≤ l. Let us set λi := (λi,1 , . . . , λi,n ). Let r > 0. We shall say that S satisfies condition (Ar ) if inf{|(Q, λ i )| = 0, 1 ≤ i ≤ l} ≥ r−|Q| .

Proposition 4.1.2 [Sto97] Let Xi = ∑lj=1 ai, j (x)S j , 1 ≤ i ≤ l be germs of holomorphic vector fields in a neighbourhood of 0 ∈ Cn , commuting pairwise. We assume

Normal form of holomorphic dynamical systems

281

that the matrix A = (ai j ) of elements ai, j ∈ On is invertible in a neighbourhood of 0 ∈ Cn . If condition (Ar ) is satisfied for some r > 0, then the vector fields Xi ’s are simultaneously and holomorphically normalizable. It is very likely that Zung’s theorem (at least in some cases) could be obtained in the following way: First of all, apply Artin’s theorem in order to transform, by an holomorphic change of coordinates, the first integrals into different monomials generating the first integrals of the Si ’s. Then, we are in situation of applying the proposition unless the matrix A is not invertible. In this last case, technics of next result could be applied. Definition 4.1.3 We shall say that s ∈ span(S1 , . . . , Sl ) is regular relatively to S whenever S
s
.1 = X .1 . X n

n

Theorem 4.1.4 [Sto05b] Let S be a diophantine family of linearly independent linear diagonal vector fields. Let X1 be a germ of holomorphic vector field with a regular linear part s at the origin. Let X2 , . . . Xl be germs of holomorphic vector fields in a neighborhood of 0 and commuting with X1 . Assume that the family {X1 , . . . , Xl } has a normal form (relatively to s) of the type l

Φˆ ∗ Xi = ∑ aˆi, j S j i=1

where the aˆi, j ’s belong to O-nS . Moreover, we assume that the family of the parts of lowest degree of this normal form is free over O-nS . Then, the family is holomorphically normalizable, i.e. there exists an holomorphic diffeomorphism in a neighborhood of 0 ∈ Cn transforming the family into a normal form. One of H. Ito results about normal forms of family of n Hamiltonian vector fields [Ito89] is a particular case the previous results. Remark 4.1.5 Contrary to Theorem 2.3.6, only X1 is required to have a nonvanishing linear part s at the origin. The linear part of the Xi ’s, i ≥ 2, could be zero. Moreover, s is not supposed to satisfy any diophantine condition.

4.2 KAM theory Let us go back to Theorem 2.3.6 and Figure 1 which refer to completely integrable systems. A natural question one may ask is the following: starting from a holomorphic singular completely integrable system in a neighborhood of the origin of Cn (a common fixed point), we consider a holomorphic perturbation (in some sense) of one its vector fields. Does this perturbation still have invariant varieties in some neighborhood of the origin? Are these varieties biholomorphic to resonant (toric)

282

L. Stolovitch

varieties? To which vector field on a resonant variety does the biholomorphism conjugate the restriction of the perturbation to an invariant variety? Is there a “big set” of surviving invariant varieties? The aim of article [Sto05a] is to answer these questions which are classical in the Hamiltonian nonsingular setting [Arn63, Arn88b, Bos86, Kol57, Ste69].

4.3 Poisson structures A Poisson structure is defined by a bracket {., .} which is applied to a couple of germs of functions to a germ of function and which satisfies the following properties : • {., .} est bilinear and skew-symmetric • { f , {g, h}} + {g, {h, f }} + {h, { f , g}} = 0 (Jacobi identity) • { f , gh} = { f , g}h + { f , h}g (Leibniz identity) It is equivalent to define a germ of two-vectors field which can be written, in a local chart, P=

∂ ∂ ∂ ∂ 1 Pi, j (x) ∧ = ∑ Pi, j (x) ∧ ∑ 2 1≤i, j≤N ∂ xi ∂ x j 1≤i< j≤N ∂ xi ∂ x j

with Pi, j = −Pj,i and which satisfies Jacobi’s identity   ∂ Pj,k ∂ Pk,i ∂ Pi, j ∑ Pi,l ∂ xl + Pj,l ∂ xl + Pk,l ∂ xl = 0 1≤l≤N for 1 ≤ i, j, k ≤ N. We want to study the singularities of analytic Poisson structures, that is points where all the Pi, j ’s vanish. First of all, we are interested in singular point where the linear part of the Poisson structure is not reduced to zero. This linear part defines on the cotangent bundle to M at 0 a structure of Lie algebra. When it is semi-simple, then J. Conn has proved that the Poisson structure is analytically conjugate to its linear part in a neighborhood of the origin [Con84]. Most of the work concerns the linearization problem (and mostly in the smooth case). We refer the book [DZ05] for these aspects and [Arn97][Appendix 14] for dimension 2. In article [Sto04] we study some analytic Poisson structures which are not even formally linearizable. Our main result is about the holomorphic normalization of some Poisson structures which associated Lie algebra is a skew-product C p  Cn . The proof uses Theorem 2.3.6 in an essential and nontrivial manner. Recently, P. Lohrmann studied analytic Poisson structures with a vanishing linear part but a nonvanishing quadratic part. J.-P. Dufour and A. Wade have defined a formal normal form for such an object [DW98]. P. Lohrmann proved the counterpart of Brjuno Theorem 2.1.1 for these Poisson structure: if the quadratic part satisfies to some diophantine condition and if the formal normal form is “completely integrable” then the Poisson structure is analytically conjugate to a normal form [Loh06, Loh05].

Normal form of holomorphic dynamical systems

283

References D. V. Anosov and V. I. Arnol d, editors. Dynamical systems. I, volume 1 of Encyclopaedia of Mathematical Sciences. Springer-Verlag, 1988. Ordinary differential equations and smooth dynamical systems, Translated from the Russian. [AGZV85] V. I. Arnol d, S. M. Guseın-Zade, and A. N. Varchenko. Singularities of differentiable maps. Vol. I, volume 82 of Monographs in Mathematics. Birkhäuser Boston Inc., 1985. [AGZV88] V. I. Arnol d, S. M. Guseın-Zade, and A. N. Varchenko. Singularities of differentiable maps. Vol. II, volume 83 of Monographs in Mathematics. Birkhäuser Boston Inc., 1988. [Arn63] V.I. Arnold. Proof of a theorem by A.N. Kolmogorov on the persistence of quasiperiodic motions under small perturbations of the hamiltonian. Russian Math. Surveys, 18(5):9–36, 1963. [Arn88a] V. I. Arnol d. Geometrical methods in the theory of ordinary differential equations, volume 250 of Grundlehren der Mathematischen Wissenschaften. Springer-Verlag, second edition, 1988. [Arn88b] V.I. Arnold, editor. Dynamical systems III, volume 28 of Encyclopaedia of Mathematical Sciences. Springer Verlag, 1988. [Arn97] V. I. Arnol d. Mathematical methods of classical mechanics, volume 60 of Graduate Texts in Mathematics. Springer-Verlag, 1997. Corrected reprint of the second (1989) edition. [Bal00] Werner Balser. Formal power series and linear systems of meromorphic ordinary differential equations. Universitext. Springer-Verlag, New York, 2000. [Bos86] J.-B. Bost. Tores invariants des systèmes dynamiques hamiltoniens (d’après Kolomogorov, Arnol’d, Moser, Rüssmann, Zehnder, Herman, Pöschel,...). In Séminaire Bourbaki, volume 133-134 of Astérisque, pages 113–157. Société Mathématiques de France, 1986. exposé 639. [Bru72] A.D. Bruno. Analytical form of differential equations. Trans. Mosc. Math. Soc, 25,131-288(1971); 26,199-239(1972), 1971-1972. [Bru95] F. Bruhat. Travaux de Sternberg. In Séminaire Bourbaki, Vol. 6, pages Exp. No. 217, 179–196. Soc. Math. France, Paris, 1995. [BS07] Boele Braaksma and Laurent Stolovitch. Small divisors and large multipliers. Ann. Inst. Fourier (Grenoble), 57(2):603–628, 2007. [Con84] J. Conn. Normal forms for analytic Poisson structures. Ann. Math., 119:577–601, 1984. [CS86] R. Cushman and J. A. Sanders. Nilpotent normal forms and representation theory of sl(2, R). In Multiparameter bifurcation theory (Arcata, Calif., 1985), volume 56 of Contemp. Math., pages 31–51. Amer. Math. Soc., 1986. [DW98] J.-P. Dufour and A. Wade. Formes normales de structures de Poisson ayant un 1-jet nul en un point. J. Geom. Phys., 26(1-2):79–96, 1998. [DZ05] J.-P. Dufour and Nguyen Tien Zung. Poisson structures and their normal forms, volume 242 of Progress in Mathematics. Birkhäuser Verlag, Basel, 2005. [Eca] J. Ecalle. Sur les fonctions résurgentes. I,II,III Publ. Math. d’Orsay. [Eca92] J. Ecalle. Singularités non abordables par la géométrie. Ann. Inst. Fourier, Grenoble, 42,1-2(1992),73-164, 1992. [Fra95] J.-P. Françoise. Géométrie analytique et systèmes dynamiques. Presses Universitaires de France, Paris, 1995. [GR71] H. Grauert and R. Remmert. Analytische Stellenalgebren. Springer-Verlag, 1971. [GY99] T. Gramchev and M. Yoshino. Rapidly convergent iteration method for simultaneous normal forms fo commuting maps. Math. Z., 231:p. 745–770, 1999. Yu. S. Il yashenko, editor. Nonlinear Stokes phenomena, volume 14 of Advances in [Il 93] Soviet Mathematics. American Mathematical Society, 1993. [AA88]

284 [Ito89] [Kol57]

[Loh05] [Loh06] [Mal82] [Mar80] [Mos90] [MR82] [MR83] [Ram93] [RS93]

[Rüs67] [Sie42] [Ste58] [Ste69] [Sto96] [Sto97] [Sto00] [Sto04] [Sto05a] [Sto05b] [Vey78] [Vey79] [Vor81] [Zun02] [Zun05]

L. Stolovitch H. Ito. Convergence of birkhoff normal forms for integrable systems. Comment. Math. Helv., 64:412–461, 1989. A.N. Kolmogorov. The general theory of dynamical systems and classical mechanics. In Proceedings of International Congress of Mathematicians (Amsterdam, 1954), volume 1, pages 315–333. North-Holland, 1957. English translation in ’Collected Works", Kluwer. P. Lohrmann. Sur la normalisation holomorphe de structures de Poisson à 1-jet nul. C. R. Math. Acad. Sci. Paris, 340(11):823–826, 2005. P. Lohrmann. Normalisation holomorphe et sectorielle de structures de Poisson. PhD thesis, Université Paul Sabatier, Toulouse, France, 2006. B. Malgrange. Travaux d’Ecalle et de Martinet-Ramis sur les systèmes dynamiques. Séminaire Bourbaki 1981-1982, exp. 582, 1982. J. Martinet. Normalisation des champs de vecteurs holomorphes. Séminaire Bourbaki, 33(1980-1981),564, 1980. J. Moser. On commuting circle mappings and simultaneous diophantine approximations. Math. Z., (205):p. 105–121, 1990. J. Martinet and J.-P. Ramis. Problèmes de modules pour des équations différentielles non linéaires du premier ordre. Publ. Math. I.H.E.S, (55):63-164, 1982. J. Martinet and J.-P. Ramis. Classification analytique des équations différentielles non linéaires résonantes du premier ordre. Ann. Sci. E.N.S, 4ème série,16,571-621, 1983. Jean-Pierre Ramis. Séries divergentes et théories asymptotiques. Bull. Soc. Math. France, 121(Panoramas et Syntheses, suppl.), 1993. J.-P. Ramis and L. Stolovitch. Divergent series and holomorphic dynamical systems, 1993. Unpublished lecture notes from J.-P. Ramis lecture at the SMS Bifurcations et orbites périodiques des champs de vecteurs, Montréal 1992. 57p. H. Rüssmann. Über die Normalform analytischer Hamiltonscher Differentialgleichungen in der Nähe einer Gleichgewichtslösung. Math. Ann., (169):55–72, 1967. C.L. Siegel. Iterations of analytic functions. Ann. Math., (43): 807-812, 1942. S. Sternberg. On the structure of local homeomorphisms of euclidean n-space. II. Amer. J. Math., (80): 623–631, 1958. S. Sternberg. Celestial Mechanics, Part II. W.A Benjamin, 1969. L. Stolovitch. Classification analytique de champs de vecteurs 1-résonnants de (BCn , 0). Asymptotic Analysis, (12): 91–143, 1996. L. Stolovitch. Forme normale de champs de vecteurs commutants. C. R. Acad. Sci. Paris Sér. I Math. (324): 665–668, 1997. L. Stolovitch. Singular complete integrabilty. Publ. Math. I.H.E.S., (91): 133–210, 2000. L. Stolovitch. Sur les structures de poisson singulières. Ergod. Th. & Synam. Sys., special volume in the memory of Michel Herman, (24):1833–1863, 2004. L. Stolovitch. A KAM phenomenon for singular holomorphic vector fields. Publ. Math. Inst. Hautes Études Sci., (102):99–165, 2005. L. Stolovitch. Normalisation holomorphe d’algèbres de type Cartan de champs de vecteurs holomorphes singuliers. Ann. of Math., (161):589–612, 2005. J. Vey. Sur certains systèmes dynamiques séparables. Am. Journal of Math. 100, pages 591–614, 1978. J. Vey. Algèbres commutatives de champs de vecteurs isochores. Bull. Soc. Math. France,107, pages 423–432, 1979. S.M. Voronin. Analytic classification of germs of conformal mappings (BC, 0) → (BC, 0) with identity linear part. Funct. An. and its Appl., 15, 1981. Nguyen Tien Zung. Convergence versus integrability in Poincaré-Dulac normal form. Math. Res. Lett., 9(2-3):217–228, 2002. Nguyen Tien Zung. Convergence versus integrability in Birkhoff normal form. Ann. of Math. (2), 161(1):141–156, 2005.

Geometric approaches to the problem of instability in Hamiltonian systems. An informal presentation Amadeu Delshams1 , Marian Gidea2 , Rafael de la Llave3 , and Tere M. Seara4

Abstract We present (informally) some geometric structures that imply instability in Hamiltonian systems. We also present some finite calculations which can establish the presence of these structures in a given near integrable systems or in systems for which good numerical information is available. We also discuss some quantitative features of the diffusion mechanisms such as time of diffusion, Hausdorff dimension of diffusing orbits, etc.

1 Introduction The goal of these lectures is to present an overview of some geometric programs to understand instability in Hamiltonian dynamical systems. Roughly speaking, the problem of instability is to decide whether the effect of small time-dependent perturbations accumulates over time. Relatedly, to show that many orbits of a time-independent Hamiltonian system explore a large fraction of the energy surface. Instability is a real problem arising in applications. For example, designers of accelerators or plasma confinement devices want to invent devices which are as devoid of instability as possible. Designers of space missions want to find orbits 1

Departament de Matemàtica Aplicada I, Universitat Politècnica de Catalunya, Diagonal 647, 08028 Barcelona, Spain e-mail: [email protected] 2

Department of Mathematics, Northeastern Illinois University, Chicago, IL 60625 e-mail: [email protected]

3

Department of Mathematics, University of Texas, Austin, TX 78712 e-mail: [email protected]

4 Departament de Matemàtica Aplicada I, Universitat Politècnica de Catalunya, Diagonal 647, 08028 Barcelona, Spain e-mail: [email protected]

W. Craig (ed.), Hamiltonian Dynamical Systems and Applications, 285–336. c 2008 Springer Science + Business Media B.V.


285

286

A. Delshams et al.

which can move freely over a wide region of space, but of course, they can only use the intertwining gravitational fields of the nearby celestial bodies. Chemists want to understand how reactions and reconfigurations of molecules take place. As is common with real problems, there are many mathematical formulations depending on the precise mathematical meaning attached to the vague words of the previous paragraph1 and many techniques which have come to bear on these formulations. For example, the lectures of Professors Cheng, Neishtadt, and Treschev in this volume present other points of view about the problem and will even present different treatments of the same mathematical model. These lectures can only aim to present informally the ideas behind some of the methods that have been proposed. We do not aim to present all the hypothesis of the results and much less complete proofs. Even when we restrict to geometric methods, we cannot aim to present a complete survey. The subject is progressing very fast. We only hope that these lectures can present an entry point to a portion the literature and indicate what to look for while reading some papers. We just want to present several milestones of the programs and to give some indication of the arguments. There are two basic steps in all the results presented here. In a first step, we will present several geometric facts that imply that there are orbits that move appreciable lengths. In a second step, we will present some finite calculations which can verify the existence of these objects in quasi-integrable systems or in systems of a special form. Hence, for some systems, deciding that instability happens can be established with a finite computation. This will have the conclusion that some types of diffusion or instability are generic in some sense in some class of systems. Remark 1.1. It should be emphasized that there are different geometric mechanisms that lead to instability. These mechanisms involve different geometric objects, have different hypothesis and lead to orbits with different characteristics. Several different mechanisms can coexist in the same model. The existence of several mechanisms was documented in some of the heuristic literature. An early paper, which is still worth reading is [LT83]. Remark 1.2. Given the practical importance of the problem of instability, there is a very large numerical and heuristic literature. Even if not easy to read, this literature contains considerable insights and can suggest several theorems. As representative papers of the numerical literature – which we cannot discuss in more detail – we mention [Chi79, Ten82, LT83, ZZN+ 89, Zas02, GLF05, FGL05, FLG06]. Perhaps the main insight from the numerical literature, is that resonances organize the diffusion (the so called Arnold web). This, indeed was one important 1 The previous paragraph contains several imprecise words such as accumulate, many, explore, large, etc. There are several rigorous formulations of these ideas. Some of the authors of this paper remember a round table in [Sim99] which included Profs. Arnol’d, Gallavotti, Galgani, Herman, Moser, Simó, Sinai. The panel was asked the question to give a canonical definition of diffusion that was preferable to the other definitions then in use. The conclusion was that it was better that each paper contains a precise definition. The reader is encouraged to compare the precise definitions of diffusion or Arnol’d diffusion used in each paper. See Remark1.3.

Geometric approaches to the problem of instability

287

motivation for several of the investigations reported here. On the other hand, we will discuss some mechanisms which do not quite fit in this paradigm. See Sections 2, 7. Remark 1.3. There are many precise mathematical formulations of what is meant by diffusion or Arnol’d diffusion. For some authors, the fact that there are whiskered tori as discussed in Section 2 is the key feature. We however take the presence large effects as the key feature. Many papers, for example [HM82] (which we will discuss more fully in Section 4) consider perturbations of size ε of an integrable system and establish existence of whiskered tori with heteroclinic intersections. These chains of whiskered tori, however are rather short and lead only to changes in the action variable of order ε 1/2 . We, on the other hand, prefer to emphasize the existence of changes of order 1 in the actions, even if they are not accomplished through whiskered tori. A careful discussion of these issues appears in [Moe96].

1.1 Two types of geometric programs With some simplification, there are two types of geometric programs that we will discuss.

Programs based on invariant objects and their relations 1. Find invariant objects (whiskered tori, normally hyperbolic invariant manifolds, periodic orbits, horseshoes, normally hyperbolic laminations, etc. as well as their stable and unstable manifolds). 2. Prove that if these objects satisfy some appropriate relations (e.g. there is a sequence of whiskered tori such that the unstable manifold of one torus intersects transversally the stable manifold of the next torus) then, there are orbits which move along the chain of invariant objects. Incipient versions of programs of this type were already present in [Poi99]. The paper which has been more influential in the mathematical literature is [Arn64, Arn63]. The main invariant objects considered in [Arn64] were whiskered tori and their invariant manifolds. We will discuss this paper in some more detail in Section 2. Other early examples of instability were [Sit53, Ale68a, Ale68b, Ale69, Ale81], which were mainly based on hyperbolic and topological properties. The study of instability properties of oscillators was pioneered in [Lit66a,Lit66b,Lev92]. Other papers establishing instability in oscillators are [AO98, Ort97, LY97, Ort04] The papers [Pus77a, Pus77b, Dou89, KPT95, Pus95] study instability in systems with collisions. The papers [Dou88, DLC83] construct examples of instability near elliptic points. The paper [CG98] revived geometric approaches and contained many useful techniques.

288

A. Delshams et al.

To study invariant objects, typically, one finds some representation of them as a function. The condition of invariance is then a functional equation, which is often studied by methods of functional analysis or just numerically or by asymptotic methods. Two very basic methods to study invariance equations are normal hyperbolicity or KAM theory. One often has to supplement them with some preliminary calculations based on averaging or on perturbative calculations.

Programs based on finite orbits “with hooks” 1. Find finite segments of pseudo-orbits such that one segment ends close to the beginning of the following segment. 2. Verify some extra properties of each of the segments. 3. Use these properties to show that there is an orbit that remains close to the whole segment of orbits. We picturesquely describe the above situation as saying that the segments of orbits have hooks so that they can be chained together. The fact that one needs some extra properties of the segments is made clear by the existence of examples – e.g. rigid rotations of the torus – where the conclusions are false. There are quite a number of mathematical results of this kind. The best known results of this type are, perhaps, the shadowing theorems for hyperbolic systems [Shu78]. The hook in this case, is hyperbolicity. For many applications, hyperbolicity is a hard hypothesis to verify – it is often even false! – so that there are many variants. See, for example [Pal00, Pil99] and references there. For us, the method which so far has proved to be more useful is the method based on correctly aligned windows. The basic idea is to use some kind of topological index of the segments of orbits so that one can show that there is an orbit in a neighborhood of the whole chain. One early example, is [Con68,Con78,Eas89]. We will discuss it in Section 5. One should also mention the variational program started in the 1930s using broken geodesics [Mor24, Hed32, Ban88]. The idea was that, if the segments are minimizers of a good variational principle, then, indeed, there are orbits that follow them.2 Some early implementations of these ideas to the problem of diffusion appear in [Bes96]. More recent applications appear in [BBB03, BCV01]. These methods also have the advantage that they apply to PDE’s [RS02, Ang87]. Very deep variational methods that also involve global considerations appeared in [Mat93, Mañ97, CI99]. Remark 1.4. Of course, there are relations between the methods. Even in [Arn64], the invariant objects were used to produce segments of orbits as well as some obstruction property which shows that there are orbits that follow the segment. In our 2

The heuristic idea is that, in the space of segments, each of our minimizers is in the center of a ball whose boundary has more action. If we take the whole orbit, the phase space is the product of the phase space of the segments so that the approximate orbit is contained in a ball so that the boundary of a ball has more action.

Geometric approaches to the problem of instability

289

discussion of applications of the method of correctly aligned windows, we will consider orbits suggested by the invariant objects.3 Even the more global variational methods of [CY04b, CY04a] start by reducing the problem using the presence of a normally hyperbolic invariant manifold. One can hope that in the near future there will be even more relations. In particular, the more local variational theories (broken geodesic methods) seem rather close to the geometric methods. One can find relations between variational methods and the windows method is [Moe05]. In these lectures, we will try to present different mechanisms as well as the verification of their presence in some quasi-integrable systems. For the geometric mechanisms we will present in these lectures, the verification of their presence in concrete systems, will involve a rather standard toolkit (averaging theory, the theory of normally hyperbolic manifolds – perturbation theory, λ -lemma –, KAM theory) and some less standard tools such as the scattering map (Section 3.2 ) and the correctly aligned windows (Section 5). We will omit most of the details, but refer to the literature. The only goal of these lectures is to present a road map to the programs and to indicate the significant mileposts to be reached. Some similar expositions are [DDLLS00, DLS03, dlL06]. The present one incorporates some progress since the previous exposition were written. Fortunately, the new developments have lead to more streamlined proofs.

2 Exposition of the Arnold example This very explicit example was constructed in [Arn64]. It is, possibly, the best known example in the mathematical literature. Some more detailed expositions of several of the aspects appear in [AA67]. A very complete explanation of the model in [Arn64] and generalizations can be found in [FM01]. In the following paragraphs, we will present the result emphasizing some of the geometric aspects that will play a role in the following. We refer [FM01] for the technical details of many of the proofs. We will emphasize several geometric properties that will play in the future. Theorem 2.1. Consider a time-dependent system defined in the action-angle variables (I, Φ ) ∈ R2 × T2 by: 1 H(I, Φ ,t) = (I12 + I22 ) + ε (cos Φ1 − 1) 2 (1) + µε (cos Φ1 − 1) (sin Φ2 + cost) , If 0 < µ  ε  1. Then, there exist orbits of the Hamilton’s equation corresponding to (1) with |I(T ) − I(0)| > 1 . 3

Strictly speaking, the windowing method only needs that they are approximately invariant.

290

A. Delshams et al.

Φ1 , I1 plane × Φ2 , I2 plane. Fig. 1 Illustration of the dynamics of the time one map of the dynamics of (1) for ε > 0, µ = 0.

We point out that the Hamiltonian (1) satisfies the conditions of KAM and Nekhoroshev theorems (in spite of being partially degenerate) [Lla01, Nie07] so that the for ε , µ small, the orbits that satisfy the conclusion occupy a small measure (these orbits cannot be in KAM tori) and T has to be very large (by Nekhoroshev’s theorem). This gives an idea of the subtleness of the phenomenon. The system (1) can be easily understood for ε > 0, µ = 0 since it is a product of two simple systems (a pendulum and a rotator), see Figure 1. We note, in particular that the manifold Λ , obtained by fixing the pendulum variables to the hyperbolic fixed point, (i.e. (I1 , Φ1 ) = (0, 0)) and letting the (I2 , Φ2 ) vary is a normally hyperbolic manifold. Clearly, Λ is topologically an annulus R × T1 . It will be important (for other mechanisms) to remark that the manifold Λ is normally hyperbolic. The main remark in [Arn64] is that the manifold Λ is foliated by invariant tori (corresponding to fixing I2 ). These tori are not normally hyperbolic (perturbations along the I2 direction do not grow exponentially), but they are whiskered tori . That is, tori, whose normal directions contain stable directions (i.e. directions which contract exponentially fast in the future) and unstable directions (i.e. directions that contract exponentially fast in the past). The rates of contraction in the future and in the past are the contracting and expanding eigenvalues of the fixed point of the pendulum. It is easy to see that they are equal to λ = ∓ε 1/2 . It is shown, in general, that to whiskered tori, one can attach invariant stable (resp. unstable) manifolds consisting of the orbits which converge exponentially fast – with a rate similar to the rate of convergence of the linearization – in the future (resp. in the past). In the uncoupled case that we are considering now, the stable and unstable manifolds can be computed explicitly. The (un)stable manifolds are just the product of the tori and the separatrix of the pendulum. In particular, the stable and unstable manifolds of a torus agree. Now, we consider 0 < µ  ε and we will treat the term containing µ as a perturbation. In such a case, we can use the general theory of whiskered tori and their manifolds. The application of the general theory to (1) is rather simple because the example has been chosen carefully so that the perturbation and its gradient vanish on Λ . Hence, the family of tori, remains the same. It is part of the general theory that the tori keep being whiskered under the new dynamics and that they have (un)stable manifolds. Furthermore, the manifolds depend smoothly on µ . The first order in the

Geometric approaches to the problem of instability

291

µ expansion can be computed easily by matching powers in formal expansions4 and it is not difficult to show that the manifolds of nearby tori intersect transversally. In some ways the result is to be expected since the µ term, even if leaving Λ invariant, is significant in the region occupied by the whiskers. It would be very easy to make perturbations with compact support intersecting the separatrices and which move them. The construction so far, for any δ > 0 allows to construct a δ pseudo-orbit that moves I2 by 1. If we start in a torus τ with an irrational rotation, we wait for the appropriate moment, then, jump in its unstable manifold, in such a way that the orbit is also in the stable manifold of another torus τ . Once we are close enough to τ , we jump into a torus with an irrational rotation – such tori are dense. Then, we can restart again. Unfortunately, this step does not allow to take the limit δ → 0 since the orbits change widely. If we make δ smaller, the orbits we constructed have to give more turns till the irrational rotation sets the phase exactly right for the jump.

2.1 The obstruction property The program of [Arn64] contains an extra step, the obstruction property – that constructs a true orbit shadowing some of the pseudo-orbits. Figure 2 depicts schematically some of the pseudoorbits made by joining heteroclinic connections and the orbits shadowing them. There is a substantial literature on the obstruction argument. We just call attention to the reader that part of the literature includes – sometimes without making it explicit – the assumption that one of the terms in the normal form of the torus vanishes. Some papers rely on normal forms to high order – hence only apply com-

Fig. 2 Illustration of some orbits in the dynamics of (1) for 0 < µ  ε . The 2 refers to the fact that Λ is two-dimensional. 4

Of course, matching powers in formal expansions does not justify that the expansions exist. In this case, using the general theory of whiskered tori, we know that these expansions exist. Historically, power matching in cases similar to this one was routinely used many years before it was justified.

292

A. Delshams et al.

fortably to C∞ or Cω systems. Others assume that all the tori can be fit in a common system of coordinates. In some papers, the construction depends on the number of tori that the orbit has to explore. Therefore, increasing the number of tori changes substantially the orbit (the time the orbit has to spend in the neighborhood of each tori increases with the total number of tori to be visited). These constructions do not allow to pass to the limit and construct orbits which visit infinitely many tori. Of course, the diversity of arguments is just a reflection of the fact that there are many types of diffusing orbits each with different quantitative and qualitative properties. We cannot survey the rather extensive literature but just call attention on some points to watch for. We certainly hope somebody will write such a survey. We also note that the obstruction argument is not the only way of constructing orbits which shadow the pseudo-orbits. In this lecture we will discuss the method of correctly aligned windows in other context, which is a topological method – applications to the shadowing of whiskered tori happen in [Rob02, GR04, CG03]. There are also variational methods [Bes96, BBB03, BCV01] for this step. In practice, the step of constructing the shadowing orbits is what controls the time T in the statement of the result. Many of the methods above lead to different estimates for T and presumably to different orbits. This again reinforces the belief that diffusion is really a superposition of several mechanisms. Here, we will just present some simple argument – we follow closely [DLS00] – which makes more precise some of the ideas in the original papers [AA67]. See also [FM01, FM03, FM00, Cre97]. The main ingredient is a somewhat sharp version of the λ -lemma – for example that in [FM00] – and a point set topology argument. Since no normal forms to higher order are used the method has only modest differentiability requirements. It can also accommodate infinitely long chains. A more elaborate argument along similar lines, but also giving more control on the orbits appears in [DLS06c]. ∞ If {Ti }∞ i=1 is a sequence of whiskered tori with irrational rotations and {εi }i=1 a sequence of strictly positive numbers, we can find a point P and an increasing sequence of numbers Ti such that

ΨTi (P) ∈ Nεi (Ti ) where Nεi (Ti ) is a neighborhood of size εi of the torus Ti . Here Ψt represents the flow associated to the system. To establish this result, note that if x ∈ WTs 1 , we can find a closed ball B1 , centered at x, and such that ΨT1 (B1 ) ⊂ Nε1 (T1 ). (2) By the λ -lemma,

/ WTs 2 ∩ B1 = 0.

Hence, there is a closed ball B2 ⊂ B1 , centered at a point in WTs 2 such that, besides satisfying (2): ΨT2 (B2 ) ⊂ Nε2 (T2 ).

Geometric approaches to the problem of instability

293

Proceeding by induction, we can find a sequence of closed balls Bi ⊂ Bi−1 ⊂ · · · ⊂ B1 ΨT j (Bi ) ⊂ Nε j (T j ), i
j. Since the closed balls are compact, they have non-empty intersection and any point in the intersection satisfies the desired property. This argument as presented above does not give estimates on the time needed to transfer. On the other hand, it gives several other information on the orbits. For example, the orbits never leave an ε neighborhood of the segments of WTs,ui so that we can be sure that the energy, or the actions, are described, up to errors of size ε by the values along the sequences visited. For future purposes, it is important to point out that the argument only uses that the tori are whiskered and it does not use at all the way that the tori fit together. Later, in Section 4, we will apply this argument to sequences of tori which are not homotopic and that, therefore, cannot be fit in common system of coordinates.

2.2 Some final remarks on the example in [Arn64] The example (1) is remarkable for many reasons. Here, we just note that the diffusion happens in places where there are no resonances. Indeed, detecting the diffusion numerically in (1) is much harder than in other examples. It is somewhat ironic that much mathematical effort was spent proving instability in models for which the result is indeed very weak. One feature of the example (1) which is important for the construction is that the second perturbation vanishes identically on a manifold. This is very non-generic and, indeed, it does not happen in many models of interest.5 We have done the first order expansion in µ , assuming ε > 0 and fixed. The dependence on ε of this theory is rather complicated. The first order term in the expansion in µ of the angle between the stable and the unstable manifoldso of a torus is of the order exp (−Aε −1/2 )µε . The remainder, on the other hand, is not easy to bound better than C µ 2 ε 2 . This is, of course, perfectly fine if µ  exp (−A/2ε −1/2 ), but if µ = ε p , then, it could happen that the leading order of the perturbation in µ does not give the whole story. As a consequence, the treatment above – based on just first order perturbation theory in µ can not establish the existence of instability in a whole ball in ε , µ or for µ = ε p .

5

One should remark, however, that it does happen in some models of interest. For example [dlLRR07] shows that perturbations which vanish on manifolds, happen naturally in some systems of physical interest such as billiards with moving boundaries and in oscillators provided that they have some symmetries and that an analysis very similar to that of [Arn64] leads to the existence of orbits of unbounded energy in these systems

294

A. Delshams et al.

3 Return to a normally hyperbolic manifold. The two dynamics approach In the exposition of [Arn64] in the previous section, we have emphasized the normally hyperbolic manifold Λ – which only appeared implicitly in [Arn64]. The reason is that the persistence of normally hyperbolic manifolds holds rather generally as was recognized in the 1960s [Sac65, Fen72, Fen74, HP70, HPS77, Pes04]. Of course, for examples other than the carefully chosen (1), one does not expect that the dynamics in the invariant manifold remains integrable. Indeed, as it is well known (we will present some ideas in Section 4.3) the resonant tori break up under perturbation so that the foliation by invariant tori gets interrupted. The general theory of normally hyperbolic invariant manifolds establishes not only the persistence of the normally hyperbolic invariant manifolds but also the existence of stable and unstable manifolds and the regularity of the dependence on parameters of these objects. A short summary of the theory of normally hyperbolic invariant manifolds can be found in Appendix A. Of course, this is no substitute for the references above. The theory of dependence with respect to parameters of normally hyperbolic invariant manifolds, justifies the perturbation theory.

3.1 The basics of the mechanism of return to a normally hyperbolic invariant manifold The basic assumption is that the stable and unstable manifolds of Λ intersect transversally. This means that there are orbits that leave the manifold but come back. We will refer to these orbits as homoclinic excursions. Note that a simple dimension counting – justified by the regularity given by the theory of normally hyperbolic invariant manifolds – shows that the set of homoclinic excursions is, locally, a manifold of the same dimension as Λ . Hence, we expect that there is an open set H− ⊂ Λ such that all the points in H− can make an arbitrarily small jump and, go into the unstable manifold of Λ , perform an homoclinic excursion and come back to Λ . Since this homoclinic excursion moves the orbit far away from Λ it is quite possible that it can be really affected by the perturbation and the action variables can change. In Section 3.2, we will describe some concrete descriptions of these sets. When the system is conservative, one expects that some of the homoclinic excursions are favorable – e.g. the excursion gains energy or action – and others are unfavorable – the excursions looses energy or actions. Since there are rather explicit formulas – which we will explain in Sections 3.2 and 4.1, one expects that the points in H− which lead to favorable or unfavorable excursions are open sets separated by a codimension 1 manifold, which can be calculated as the zero set of a function (in the models discussed in Section 4 perturbative formulas for this function are rather standard).

Geometric approaches to the problem of instability

295

Fig. 3 Illustration of orbits that gain energy by intertwining homoclinic excursions with staying around an invariant manifold.

Note that H− and the separation between the favorable and unfavorable regions depend very strongly on the perturbation far away from Λ . Hence, we can expect that the dynamics on Λ – which is unaffected by the perturbations away from the manifold – is completely unrelated to the separation between favorable and unfavorable excursions. Hence, unless this separation is invariant under the dynamics in Λ , one can stay around Λ for a carefully chosen time and move into the favorable region. We emphasize that, explicit perturbative computations can give approximations to the manifold separating the favorable from the unfavorable excursions, so that a finite computation can establish that there are orbits in Λ that move into the favorable region. In this way, for many systems, one can construct pseudo orbits by interleaving orbits that follow a homoclinic excursion and orbits that remain in Λ so that we go from the end of a homoclinic excursion to another favorable excursion. This can be compared to primitive sailing: When the wind is favorable, the boat moves. When the wind turns bad, it moves to the coast and anchors waiting till the wind becomes favorable again. Of course, if one is interested in true orbits rather than on δ pseudo-orbits with δ arbitrarily small, one still needs an extra step – shadowing or obstruction. Some versions of these arguments are discussed in Sections 2.1 and 5. To make the above heuristic ideas rigorous, one uses: (a) a tool to describe the homoclinic excursions, which allows explicit computations (b) some explicit description of the dynamics on Λ , (c) some tools to pass from the pseudo-orbits to the orbits. Of course, the analysis of the dynamics restricted to Λ is just the general problem of dynamical systems. The description of homoclinic intersections will be undertaken in Section 3.2. We note that the scattering map is not the only way to discuss homoclinic excursions. The paper [Tre02a, Tre02b] introduce the separatrix map. We also call attention to [BK05].

296

A. Delshams et al.

3.2 The scattering map The scattering map is a particularly convenient way of describing the behavior on a homoclinic excursion. It was introduced explicitly in [DLS00] as a geometrically natural alternative to Melnikov theory so that issues of domain and monodromy could be discussed in detail. Related ideas for center manifolds were introduced in [Gar00]. A much more systematic theory of the scattering map was developed in [DLS06a]. An orbit is homoclinic if the future and the past converge exponentially fast to Λ . We adopt the same notation as in Appendix A. We recall that the stable and unstable manifolds can be decomposed into stable ' ' manifolds of single points, namely: WΛs = x∈Λ Wxs , WΛu = x∈Λ Wxu . The above / decompositions are are foliations because if x, y ∈ Λ , x = y, then Wxs ∩ Wys = 0, / We will refer to these foliations as Fs,u respectively. Wxu ∩Wyu = 0. Using the foliations Fs,u we can define the “wave operators” Ω+ , Ω−

Ω± : WΛs,u −→ Λ defined by

x ∈ WΩs + (x)

x ∈ WΩu − (x)

(1) (2)

If there is a manifold Γ such that Ω− is a diffeomorphism from Γ to Γ )−1 : H Γ → Γ and relatedly, its range Ω− (Γ ) ≡ H−Γ , then we can define (Ω− − ⊂ WΛs

∩WΛu

Γ −1 sΓ = Ω+ ◦ (Ω− )

(3)

See Figure 4. This set H−Γ is the set of initial points of trajectories having the property that an small push can make them go through Γ . This is a more precise version of the set H− wich we discussed in Section 3.1. The set H−Γ specifies that the connections go through Γ . The map sΓ : H− → H+ , gives an encoding of the homoclinic excursions that pass through Γ . If we consider one such excursion, the orbit is asymptotically close to one orbit in Λ in the past and to another orbit in Λ in the future. The map sΓ gives the future orbit as a function of the asymptotic orbit in the past.6 Of course, the scattering map depends very strongly on the manifold Γ we have chosen. Escaping from Λ along different routes will, clearly, have very different effects and the scattering map will be very different. Some examples of celestial mechanics with explicit computations appear in [CDMR06]. Γ is invertible from Now, we discuss some natural hypothesis that imply that Ω− its range to Γ and that this is maintained under perturbations and that there is good dependence with respect to parameters. Basically, we will reduce the definitions to 6

This is remarkably similar to the definition of the scattering matrix in the time-dependent scattering theory in quantum mechanics. Indeed, there are many more analogies and we have chosen the notation to reflect them. Other classical analogues of quantum scattering theory, somewhat different from those considered here, were considered in [Hun68, BT79, Thi83] and in a more general context in [Nel69].

Geometric approaches to the problem of instability

297

Fig. 4 Illustration of the definition of the scattering map.

transversality conditions so that the implicit function theorem gives the persistence and smooth dependence on parameters. A natural set of conditions to define scattering map is that for all x ∈ Γ , TxWΛs + TxWΛu = Tx M TxWΛs ∩ TxWΛu = TxΓ TxWΩs + x ⊕ TxΓ = TxWΛs TxWΩu − x ⊕ TxΓ = TxWΛu

(4)

(5)

The conditions in (4) mean that WΛs , WΛu “intersect transversally” along Γ . The first condition in (5) means that Γ is “transversal to the foliation” {Wxs }x∈Λ inside WΛs . The second equation in (5) means that Γ satisfies an analogous property relative to the unstable foliations. See Figure 5. If we have (4) for just one x0 , the implicit function theorem tells us that we can find a smooth manifold Γ containing x0 such that (4) holds for all x ∈ Γ . Since the manifold Γ is obtained applying the implicit function theorem, if both WΛs , WΛu , are Ck manifolds in a neighborhood of x, then Γ will also be a Ck manifold. Similarly, applying the the implicit function theorem, the regularity theory for the manifolds and their smooth dependence on parameters, discussed in Appendix A, we conclude that if fε is a C1 family and f0 has a Λ0 , Γ0 satisfying the normal hyperbolicity and transversality conditions, that there is a C1 family of manifolds Λε which are normally hyperbolic and another family of manifolds Γε satisfying the are C−1 families, we obtain properties. In the case that we can guarantee that WΛs,u ε −1 that Γε is a C family and we can also obtain smooth dependence on parameters Γε and for the scattering map.7 for the Ω± 7

The smooth dependence of a map in domains which are changing, should be understood in the sense that there is smooth family of maps from a fixed domain to the domains so that the composed map is smooth.

298

A. Delshams et al.

Fig. 5 Illustration of the conditions in (5).

The properties in (5) are very different. Even if the formulation of (5) does not require that the foliations Fs,u are smooth, they become more interesting when these foliations are C1 foliations. In this case, the implicit function theorem tells that, when we move along Γ , we have to move across the foliation. The implicit function theorem shows that, if the foliations Fs,u are C1 – this is implied by properties of the hyperbolicity constants, so that it holds true in some C1 open sets of examples – and (5) hold, then, Ω± are locally invertible. Again, because this is just an application of the implicit function theorem and there is a good dependence on parameters, we obtain if (4), (5) are satisfied for a map, they will be satisfied – with a similar Λ , Γ – for all the small C1 perturbations. Furthermore, if we consider smooth families of maps, there will be smooth dependence on parameters. Remark 3.1. One could argue heuritically that (5) could fail in a codimension 1 set of Γ – transversality is a codimension 1 phenomenon. Of course, this heuristic argument, could be false. Notably, the heuristic argument is false for the models considered in [DLS00, DLS06c]. It however, applies to some examples considered in [DLS06b]. Nevertheless, as shown in [DLS06b], the existence of an open set is enough for the construction of orbits that move appreciable amounts. One can also note that one expects to have infinitely many Γ , each of which with a different scattering map. The argument does not require that all the excursions go through the same Γ , so that the set of points which cannot be moved by this argument should be empty in manu examples.

3.3 The scattering map and homoclinic intersections of submanifolds One important application of the scattering map is that it allows us to discuss transversal intersections of WΣu1 , WΣs2 where Σ1 , Σ2 ⊂ Λ are invariant manifolds

Geometric approaches to the problem of instability

299

under the map f . One example is, of course, the whiskered tori inside the manifold Λ that were discussed in Section 2. In Section 4 we will see other examples that are more challenging. It was shown in [DLS00, DLS06b, DLS06c] that if, for some manifold Γ , satisfying (4) (5), we have8 sΓ (Σ1 ) Λ Σ2 . (6) Then, WΣu1  WΣs2 .

(7)

This result is useful because the hypothesis (6) can be verified by calculations on the invariant manifold Λ . The conclusions is that the (un)stable manifolds of σ1 , σ2 are transverse in the full manifold M. In the case that Σ1 , Σ2 are invariant circles which are close together, the transversality of intersections is usually discussed using Melnikov theory. Notice, however that Melnikov theory – since it is based on first order calculations often done in a concrete coordinate system – requires that the manifolds Σ1,2 are expressed in the same system of coordinates, in particular, they are homotopically equivalent. The above result, however, is coordinate independent. This is crucial for the applications in [DLS06b], discussed in Section 4, where Σ1,2 are not topologically equivalent. As we will see in Section 3.7 there are rather explicit – rapidly convergent – formulas for the perturbative computation of the scattering map. Therefore, the theory outlined above can give rather efficient ways of establishing intersections.

3.4 Monodromy of the scattering map Γ are locally invertible, they could fail to be invertible in a domain which Even if Ω± is large enough to include non-contractible closed loops. One interesting example was discussed already in [DLS00] and, in more detail in [DLS06c, DLS06a]. For example, when considering stable manifolds of a periodic orbit λ , the intersection manifold Γ looks like a helix. That is, if we increase the phase of the intersection, then, eventually we go into a different homoclinic intersection of the time-1 map. This is a geometric counterpart of the fact that, in some calculation in first order perturbation theory of intersections of invariant manifolds – often called Melnikov theory – one has to add real variables to angle variables. See Figure 6.

3.5 Smoothness and smooth dependence on parameters Note that the sufficient conditions (4), (5) that ensure the existence of the scattering map in a neighborhood are transversality conditions that are robust under We use the notation Λ to indicate that the manifolds intersect transversally as manifolds in Λ . In particular, when we use this symbol, we assume that the intersection is not empty.

8

300

A. Delshams et al.

Fig. 6 Illustration of the monodromy of the scattering map for the stable manifolds of periodic orbits.

perturbations. Hence, given a concrete system, they can be established with a finite precision calculation. Later, in Section 4.1 we will see how they can be verified by perturbative calculations from an integrable system. See [DLS06b, GL06a]. The conditions can also be verified numerically if one controls the precision of the calculations [CDMR06]. It follows from the general theory of dependence on parameters that, under the conditions (4), (5), and smoothness of the foliations Fs,u then, the scattering map is smooth jointly on the manifold and on parameters.9

3.6 Geometric properties of the scattering map So far, the discussion of the scattering map has only used normal hyperbolicity and regularity of the maps considered. If the maps fε have some geometric structure, the scattering map also inherits some geometric properties. Notably, if fε is symplectic (resp. exact symplectic) and Λ0 is a symplectic manifold (hence, exact symplectic if fε is exact symplectic) then sε is a symplectic (resp. exact symplectic) family of maps. This was proved in [DLS06a]. In the context of center manifolds it was proved in [Gar00].

9

The discussion of smoothness with respect to parameters of the scattering map presents some technical annoyances such as that the domain of sε is Λε , which changes as ε changes. An easy solution is to consider smooth (jointly with respect to the coordinates and the parameters) parameterizations kε of the invariant manifold Λε . That is kε (Λ0 ) = Λε . See Appendix A.

Geometric approaches to the problem of instability

301

There are two important consequences of the symplectic character. • There are many techniques to discuss intersections of Lagrangian manifolds under symplectic mappings, see [Wei73, Wei79]. • There are very efficient perturbation theories for symplectic mappings. Historically this one of the reasons why Hamiltonian formalism was invented. We will discuss several versions of Hamiltonian perturbation theory here. Taking advantage of both features at the same time, one gets a very efficient perturbative theory for the intersections of manifolds under the scattering map. In view of the results mentioned in Section 3.3, this is very useful to obtain transition tori. In [DLS06a] it was proved that there is a natural smooth parameterization kε (Λ0 ) = Λε such that k0 is the immersion and that kε∗ ω – the pull–back by kε of the symplectic form ω – is independent of ε . This later condition is a natural normalization and it is shown in [DLS06a] that this natural normalization determines uniquely the deformation. Then, denoting by sε the scattering maps generated by a smooth family of maniΓ , we have that folds Γε satisfying (4), (5), and invertibility of Ω− s˜ε ≡ kε−1 ◦ sε ◦ kε

(8)

is symplectic under kε∗ ω ≡ k0∗ ω . Note that s˜ε : Λ0 → Λ0 can be thought of as the expression of sε in the coordinates kε mentioned above. Furthermore, in [DLS06a], one can find explicit perturbative formulas for the canonical perturbation theory of s˜ε . We will summarize them in Section 3.7.

3.7 Calculation of the scattering map Given families of exact symplectic mappings there are very efficient ways of computing perturbation theories using the deformation method of singularity theory [LMM86]. If gε is a family of exact symplectic mappings, it is natural to study instead the vector field Gε generating the family. d gε = Gε ◦ gε . dε

(9)

The fact that gε is exact symplectic for all ε is equivalent to g0 being exact symplectic and ıGε ω = dGε (here ıGε ω is the contraction of vectors and forms). Under enough regularity conditions, (9) admits a unique solution. Hence, it is the same to work with Gε or Gε . The interesting thing is that the family of functions Gε satisfies much simpler equations. The reason is that the Gε – and hence Gε can be thought as infinitesimal deformations and the only equations that one can form with infinitesimal quantities are linear.

302

A. Delshams et al.

In the following, we will apply this idea to gε being several of the families appearing in the problem. We will keep the convention of keeping the same letter for the objects corresponding to a family. We will use caligraphic for the vector field and capitals for the Hamiltonian. In [DLS06a], it is shown that there are remarkably simple formulas for S˜ε , the generator of the the map s˜ε – the expression of sε in coordinates. N− −1

S˜ε = lim

∑

N± →+∞

−j −1 ε −1 Fε ◦ fε− j ◦ (ΩεΓ− ) ◦ s−1 ε ◦ kε − Fε ◦ f ε ◦ sε ◦ kε

j=0

N+

ε −1 + ∑ Fε ◦ fεj ◦ (ΩεΓ+ ) ◦ kε − Fε ◦ fεj ◦ kε

j=1

N− −1

= lim

N± →+∞

∑

(10) Fε ◦

ε −1 fε− j ◦ (ΩεΓ− ) ◦ kε

◦ s−1 ε − Fε

◦ kε ◦ rε− j ◦ s−1 ε

j=0

N+

ε −1 + ∑ Fε ◦ fεj ◦ (ΩεΓ+ ) ◦ kε − Fε ◦ kε ◦ rεj

j=1

Similarly, for Hamiltonian flows, we have  0

Sε = lim

T± →∞ −T−

dHε ε −1 ◦ Φu,ε ◦ (ΩεΓ− ) ◦ (sε )−1 ◦ kε dε

dHε ◦ Φu,ε ◦ (sε )−1 ◦ kε dε  T+ dHε dHε ε −1 + ◦ Φu,ε ◦ (ΩεΓ+ ) ◦ kε − ◦ Φu,ε ◦ kε dε dε 0 −

(11)

It is not difficult to see that the sums or the integrals converge uniformly. The formulas (10) and (11) give the hamiltonian of the deformation as the integral of the generator of the perturbation over the homoclinic orbit minus the generator of the perturbation evaluated on the asymptotic orbits. Note that, because of the exponential convergence of the homoclinic orbits and their asymptotic orbits, it is not difficult to see that the integrals in (10) and (11) converge exponentially fast. In [DLS06a] one can also find that derivatives up to an order (which is given by ratios of convergence exponents) also converge exponentially fast. The effect of the homoclinic excursions on slowly changing variables can be computed using more conventional methods – we will present some of these computations in Section 4.1. One novelty of the geometric theory presented in this section is that it allows computation of the effect of the homoclinic excursions not only on the slow variables, but also on the fast variables. Notice also that, we can compute the intersection between objects of different topologies very simply. This extends many calculations usually done using

Geometric approaches to the problem of instability

303

Melnikov theory. It suffices to apply (6). Note that the present theory only involves convergent integrals. This was somewhat controversial in the so-called Melnikov theory. See [Rob88].10 The Hamiltonian theory is particularly effective when the manifolds Σ are level sets of a function. We will see some examples in Section 4.6.

4 The large gap model The model is basically a rotor coupled to one or several penduli and subject to a periodic perturbation. This model was introduced in [HM82], but it appears naturally as a model of the motion near a multiplicity 1 resonance. A fuller treatment of multiplicity 1 resonances appears in [DLS07]. One could consider that it is a version of the example (1) when we set ε = 1 (hence rename as ε the parameter µ in (1)) but we allow the perturbing term to be a general one. In the paper [GL06b] it was remarked that the fact that the pendulum variables have only 1 degree of freedom can be easily removed and one could consider many penduli. Hence, the geometric treatment can be easily generalized to the case that the hyperbolic variables have several components. Hence, we consider the model   n 1 2 p +Vi (qi ) + h0 (I) Hε (p1 , . . . , pn , q1 , . . . , qn , I, φ ,t) = ∑ ± 2 i (1) i=1 + ε h(p1 , . . . , pn , q1 , . . . , qn , I, φ ,t; ε ), where (pi , qi ), (I, φ ) are symplectically conjugate. We will assume that Vi (0) = 0, Vi

(0) > 0. This means that Vi has a non-degenerate local minimum – that we set at 0. We will also assume that the pendulum Pi has a homoclinic orbit to 0. This is implied by the fact that there is no other critical point p with Vi (p) = 0. Both conditions are implied by Vi being a Morse function. The version of (1) considered in [HM82,DLS06b,GL06a] consider only the case n = 1, but, as we will see, the complications introduced by several variables is not too important. A full treatment of (1) for general n appears in [GL06b]. We will explain it in Section 4.1. One extra assumption in [DLS06b] – which we will maintain in the discussion in this section – is that the perturbation term h was a trigonometric polynomial in the angle variables. This assumption simplifies the calculations since there is only a 10

Unfortunately, many references in Melnikov theory still invoke the use of Melnikov functions given by integrals of quasi-periodic functions. The textbook explanation is that these integrals converge along subsequences. Unfortunately, the resulting limit – and hence the predictions of these theories – depend on the sequence taken, so that the textbook explanation cannot be true. The real explanation is that these references forgot to take into account some important effect. In many cases, it is the change of the target manifold.

304

A. Delshams et al.

finite number of resonances to be studied. It allows us to emphasize the geometric objects appearing at each resonance. When h is not a polynomial, for each value of ε > 0 it suffices to study a finite number of resonances, but the number of resonances to be considered is ε −α . One needs to do some rather explicit quantitative estimates on the resonances. The assumption that the perturbation is a polynomial has been removed by very different methods. The paper [DH06] contains a very deep study of resonances taking into account the effect of the size of the Fourier coefficients on the size of the resonant region. The paper [GL06b] considers very large windows, much larger than the resonance zones and uses the method of correctly aligned windows to conclude existence of diffusion without having to analyze what happens in the region of resonance. This leads to less conditions than the analysis in [DLS06b, GL06a]. Also, the method in [GL06b] leads to optimal estimates on the time. The analysis of (1) we will present starts by noting that Λ0 = {pi = 0, qi = 0} is a normally hyperbolic invariant manifold for the time-1 map. Applying the theory of normally hyperbolic manifolds, we conclude that, for ε small enough, it persists. In contrast with the example (1), the motion on the invariant manifold will not remain integrable. Indeed, the foliation of KAM tori will present gaps of size ≈ ε 1/2 . In the rest of the section, we will describe how to construct orbits that indeed jump over the resonance zone.11

4.1 Generation of intersections. Melnikov theory for normally hyperbolic manifolds In the model (1), even if the manifold Λ0 is normally hyperbolic, its stable and unstable manifolds coincide. In this section, we want to argue that, under some non-degeneracy conditions on h which we will make explicit, for 0 < |ε |, there is a manifold Γε satisfying the conditions (4), (5). Furthermore, one can define the scattering map in a patch which is rather large and uniform with respect to ε . The fact that there is a Γε which depends smoothly on parameters and, in particular, can be continued through ε = 0 is well known to experts and we present the ideas of a simple proof later. See also [GL06b]. These are sometimes called primary intersections of the stable and unstable manifolds, to distinguish them from other intersections which do not have a limit as ε → 0. See [Mos73, p. 99 ff.]. Subsequent steps of the construction of diffusing orbits could use any of these intersections for which the next non-degeneracy assumptions can be verified. The calculations we will develop here will work just as well for any of the primary intersections. The use of the secondary intersections deserves more study. 11 The paper [HM82] showed only that there were heteroclinic intersections between some whiskered tori. The length of the heteroclinic chains constructed in [HM82] goes to 0 as ε → 0. This was the meaning of Arnol’d diffusion adopted in that paper. It is very interesting to compare the Melnikov theory developed there with the based on the scattering map.

Geometric approaches to the problem of instability

305

Very elegant geometric theories of intersections of stable and unstable manifolds can be found in [LMS03]. In these lectures, we will follow [GL06b] and present a very simpleminded calculation for the model using coordinates. The paper [GL06b] contains significantly more details than those presented here. We call attention that the calculation here does not assume that the variables I, φ in (1) are one-dimensional. This will play a role in Section 7. A key observation is that, by the theory of normally hyperbolic manifolds, we aldepend smoothly on parameters. We just need to compute ready know that Λε , WΛs,u ε explicitly what are the derivatives of these objects. The non-degeneracy conditions alluded above are just that the first order in ε calculation predicts an intersection satisfying (4), (5). If the first order perturbation predicts a transversal intersection, the implicit function theorem allows us to conclude that indeed there is an intersection, and that the formal calculation gives the leading order. For this calculation, the fundamental theorem of calculus will play an important role, hence it is better to consider flows rather than time-1 maps. To make it autonomous, we will just add a variable t. We will use the notation Λ˜ to refer to the invariant manifold in these coordinates. For each of the penduli, we choose a homoclinic orbit xi and consider the unperturbed homoclinic manifold {(x1 (τ1 ), x2 (τ2 ), . . . , xn (τn ))}.12 The variables τi are variables parameterizing the separatrix of the i pendulum. We note that in a neighborhood of the homoclinic manifold – excluding a neighborhood of the critical points – we can extend the variables τi . The variables τi and Pi constitute a good system of coordinates in this neighborhood. See Figure 7 for an illustration of this system of coordinates. Again, appealing to the smoothness of the dependence of the stable manifolds on parameters, we know that the perturbed manifolds can be written as the graph of a function that gives the Pi as a function of τ , I, φ ,t. Furthermore, this function will depend smoothly on parameters. Our only goal, then, is to compute the first order expansion of this function, knowing already that such an expansion exists. We will denote the time evolution of a point by Ψεs . Remember that, to make the system autonomous, we consider t as a variable, which takes values on a circle. We will denote the invariant manifolds in the extended phase space as Λ˜ . Let x be a point in WΛs˜ , by the fundamental theorem of calculus, we have, for ε any T , Pi (x) − Pi (Ω+ε x) = Pi (ΨεT (x)) − Pi (ΨεT (Ω+ε x)  T  d  Pi (Ψεs (x)) − Pi (Ψεs (Ω+ε x)) ds − 0 ds

12

Note that, in general, each of the penduli will have two homoclinic orbits to the critical point (one going in one direction and the other going in the opposite direction). So that, there will be 2n homoclinic manifolds with parameterizations similar to the ones considered in the text. Since the conditions we will considering be are sufficient conditions for existence of unstable orbits, having many orbits at our disposal makes it more likely that we have instability.

306

A. Delshams et al. P xuε Λε Λε

xsε

τ

Λε

ε

Fig. 7 Illustration of the system of coordinates in a neighborhood of the homoclinic manifold.

and, taking limits T → ∞, we obtain Pi (x) − Pi (Ω+ε x) = −

 ∞ d  0

ds

 Pi (Ψεs (x)) − Pi (Ψεs (Ω+ε x)) ds

(2)

Now, recalling that we are only computing up to order ε , we can simplify significantly the formula. We note that because Pi has a critical point at 0, we have Pi (Ω+ε x) = O(ε 2 ), We also note that    d  Pi (Ψεs (x)) − Pi (Ψεs (Ω+ε x) = ε {Pi , h} ◦ Ψεs (x)) − {Pi , h} ◦ Ψεs (Ω+ε (x))) ds = O(ε ) where {·, ·} is the Poisson bracket. Notice also that the integrand in (2) is converging exponentially fast to zero. Hence, we have: Pi (x) = −ε

 c| log(ε )| 0

  ds, {Pi , h}(Ψεs (x)) − {Pi , h}(Ψεs (Ω+ε x)) + O(ε 2 )

Since the integral is over a finite interval, we observe that, if |s|
c| ln(ε )|, then |Ψεs (x) − Ψ0s (x)|
c| ln(ε )|ε Also, using the smooth dependence of the stable and unstable foliations, we obtain that |Ψεs (Ω+x ) − Ψ0s (Ω+0 x)|
c| ln(ε )|ε

Geometric approaches to the problem of instability

307

Hence, we can transform the integral into Pi (x) = −ε

 c| log(ε )| 0

  ds, {Pi , h}(Ψ0s (x)) − {Pi , h}(Ψ0s (Ω+0 x)) + O(ε 2 | ln(ε )|)

Remark 4.1. The above calculation identifies the derivative of the manifold with respect to ε when we consider the C0 topology of functions. In the case that we know that the derivative in Cr sense exists, the previous expression has to be the derivative in the Cr sense too. In [GL06b], one can find justification of the slightly stronger result that the integrals above converge uniformly in Cr – provided that the Hamiltonians are uniformly Cr+2 . A very similar formula – reversing the time – can be obtained for an expression of the unstable manifold as a graph. Subtracting them, we obtain an expression for the first order expansion of the separation ∆ of the Pi coordinates of the manifolds as a function of the τi , I, φ ,t

∆i (τ , I, φ ,t; ε ) = ε∆i0 (τ , I, φ ,t) + O(ε 2 ) where the O(ε 2 ) can be understood in the sense that the C1 norm is bounded by Cε 2 . The implicit function theorem shows that if we find a zero of ∆i0 = 0 which is non-degenerate (i.e., rank Dτ ∆ 0 = n) then we can find τ ∗ (ε , I, φ ,t) such that ∆ (τ ∗ (ε , I, φ ,t), I, φ ,t; ε ) = 0. Hence, substituting in the variables P we can onbtain a parameterization of the intersection. A more detailed analysis shows that the expressions of ∆i are derivaties of a potential function with some periodicities [DR97]. Hence they have to have zeros. The assumption that these zeros are non-degenerate is a mild non-degeneracy assumption that can be verified in practical problems. It also holds generically. The case n = 1 is studied in great detail in [DLS06b]. In [GL06b] one can find an study of how to produce several of these solutions for n > 1.

4.2 Computation of the scattering map The calculation of the scattering map in this case can be done as a particular case of the general theory of Section 3.2. Notice that the formulas (10) are given in terms of limit of the intersection as ε → 0, which we computed in the previous section using the easy part of the Melnikov theory. The calculation in [DLS06b], was done by a different method since at the time that [DLS06b] was written, the authors were not aware of the symplectic theory of the scattering map. The method of [DLS06b] was more elementary. Only the effect of the scattering map in one of the coordinates was computed. This was done using the fact that one

308

A. Delshams et al.

of the coordinates in the invariant manifold – namely the energy – has a slow variation, so that in the calculation of the change of energy along a homoclinic excursion, one can use – up to the accuracy needed – just the fundamental theorem of calculus integrating over the unperturbed trajectory. The calculation can be done in very similar way to the calculation done in Section 4.1.13 The fact that in [DLS06b] one only got control on one of the variables made the calculation of subsequent properties more complicated than what is nowadays possible using geometric theory. See [DLS07]. On the other hand, the calculation based on estimating the change of energy is natural for the purposes of the study of the intersection with KAM tori – which are given as level sets of the averaged energy. For the purpose of this exposition, we will just mention that, for the model considered, once we settle on one primary homoclinic intersection, the scattering map can be computed as an explicit perturbation series with well controlled remainders. As in all the steps of this strategy, the calculations required can be done by very different methods. The more modern methods, taking more advantage of geometric cancellations seem more efficient even if the older methods can compute some features faster. The conclusions is that – under conditions which can be checked explicitly and which, in particular, hold generically – the domain of definition of the scattering map contains a set which is independent of ε as ε → 0. We call attention to the fact that the formulas for the scattering map depend heavily on the behavior of the perturbation along the whole homoclinic excursion.

4.3 The averaging method. Resonant averaging The averaging method for nearly integrable systems goes back at least to [LP66]. Modern expositions are [LM88, AKN88, DG96]. An introduction for practitioners is [Car81]. See also [Mey91]. The basic idea is very simple. Given a quasi-integrable system, one tries to make changes of variables that reduce the perturbed system to another integrable system up to high powers in the perturbation parameter. This is accomplished by solving recursively cohomology equations. There are many contexts and variations which make the literature extensive, even if there is only one guiding principle. For example one can consider autonomous perturbations or periodic perturbations, maps, flows etc. There are different possible meanings of “as simple as possible”. One difference that leads to several variants is the fact that one can parameterize perturbations in different ways (generating functions, several types of Lie Series, deformation method, etc.) A systematic comparison of differences between these perturbation theories was undertaken in [LMM86]. 13 The actual calculation done in [DLS06b] uses not the energy – which is easily seen to be an slow variable – but rather a linear approximation to the energy. This makes only higher order differences. This linear approximation had been used customarily in the literature. At the time that [DLS06b] was written, it was important to make contact with the previous literature.

Geometric approaches to the problem of instability

309

In the present problem, we consider periodic perturbations of integrable flows with one degree of freedom. To make comparisons with the literature easier, it will be convenient to make the system autonomous and symplectic by adding an extra variable A symplectically conjugated to t Hε (I, φ ,t, A) = H 0 (I) + A + ε H 1 (I, φ ,t) + ε 2 H 2 (I, φ ,t) + · · ·

(3)

where, of course, H(I, φ ,t + 1, A) = H(I, φ ,t, A), so that t can be considered as an angle variable. The A is added to keep the symplectic structure. Notice that it does not enter into the evolution of the other variables. Again, for the sake of expediency in this presentation, we will omit considerations of issues of differentiability, estimates of reminders, etc. We refer to [DLS06b, Section 8], but the averaging method is covered in many other references, including some of the lectures in this volume. For simplicity also, we will assume that all the terms in the expansion in ε are trigonometric polynomials with the same set of indices. That is, H i (I, φ ,t) =

∑

i Hk,l (I) exp(kφ + lt).

(4)

k,l∈Ni ⊂Z2

Note that in the Appendix A, we show that this assumption for the case that we are interested in, follows from the assumption that the h in (1) is a trigonometric polynomial. The general theory of averaging does not require this assumption, but it involves several analysis consideration, which we prefer to avoid in an exposition. We try to find a time periodic family of symplectic changes of variables kε (I, φ ,t) = (I, φ ) + O(ε ) in such a way that Hε (kε (I, φ ,t),t) is as simple as possible. One possible way to try to generate the kε ’s is to write them as the time-1 solutions of a differential equation d s k = ε J∇Kε ◦ kεs , ds ε

kε0 = Id,

where J is the symplectic matrix. In this case, we consider the evolution in the p, q, A, I, φ ,t variables and the ε is just a parameter (this is not what we did in the section on deformation method). The gradient ∇ refers to the p, q, A, I, φ ,t variables. The function Kε is called the Hamiltonian. This way of parameterizing changes of variables is one of the variants of Lie transforms, [Car81, Mey91]. We will assume that Kε = ε K 1 + ε 2 K 2 + · · · , It is well known from Hamiltonian mechanics [Arn89,AM78,Car81,Mey91] that Hε ◦ kε = H 0 + ε (H 1 + {H 0 + A, K 1 }) + O(ε 2 ) where {·, ·} denotes the Poisson bracket in the variables I, φ , A,t. Therefore, our goal is to find K 1 in such a way that R1 ≡ H 1 + {H 0 + A, K 1 }

(5)

310

A. Delshams et al.

is somewhat simple (we will make precise what “simple” means in our case). Since R1 is the dominant term in Hε ◦ kε , one can hope that the dynamics expressed in the new coordinates is simple. In terms of Fourier coefficients, (5) is equivalent to 1 1 (I) + i(kω (I) + l)Kk,l (I), R1k,l (I) = Hk,l

(6)

where ω (I) = ∂∂I H0 (I). The assumptions include that H0 is twist. That is that ω (I) is monotonic, so that for each k, l there is one and only one pk,l such that kω (Ik,l ) + l = 0. Of course, Ink,nl = In,k . The points Ik,l are called resonances. Because of the assumption that the perturbation is a polynomial, we have to consider k, l ranging only over the finite set N ⊂ Z2 . We see that (6) has very different character depending on whether (kω (I)+l) = 0 i (I) or not. If (kω (I) + l) = 0, we have to set R1j,k (I) = H 1j,k (I) but we can choose Kk,l arbitrary. Since we want that our solutions are differentiable, we have to make sure that the choices are made in a differentiable way. A particularly simple way – used in [DLS06b] to make these choices is to take a fixed C∞ cut-off function Ψ and a fixed number L so that denoting ΨL (t) = Ψ (t/L), we take the choice R1 (I, φ ,t) =

∑

1 ΨL (I − Ik,l )Hk,l (Ik,l ) exp(i(kφ + lt)),

∑

1 (1 − ΨL (I − Ik,l ))/i(ω (I)k + l)Hk,l (I) exp(i(kφ + lt).

k,l∈N

K (I, φ ,t) = 1

(7)

k,l∈N

If we choose conveniently L – we are considering only a finite number of resonances – we can ensure that the intervals [−2L + Ik,l , 2L + Ik,l ] do not intersect for different resonances. So, we can divide the phase space into two regions: • One “non-resonant region” where – in the appropriate coordinates – the system is integrable up to an error of order ε 2 . • A finite number of “resonant regions”. Each of the resonant regions can be labeled by a frequency l/k expressed in an irreducible fraction. In one of these resonant regions, in the appropriate coordinates, the Hamiltonian is14 : 1 H0 (I)+A + ε ∑ Hnk,nl (Ik,l ) exp(in(kφ + lt)) + O(ε 2 ) n∈

=

(8)

H0 (I) + A + ε V (kφ + lt) + O(ε 2 ).

The dynamics of the Hamiltonian (8) are easy to understand. If we introduce the variables φ˜ = kφ + lt, I˜ = I − Ik,l – this change of variables is not symplectic, but it just multiplies the symplectic structure by a constant, so that the equations of motion – up to a constant change in time are also given by a Hamiltonian. Note also 14

Again, we ignore regularity issues. It is not hard to show that if we assume that the function H 1 is Cr , then, K 1 , R1 are Cr−2 so that the error term in (8) can be considered in the Cr−2 norm. Again, we refer to [DLS06b].

Geometric approaches to the problem of instability

311

that in this change of variables, the period of the angle variables is changed. Hence, in the new variables, the Hamiltonian is: ˜ + A + ε Vk,l (φ˜ ) + O(ε 2 ) α Hk,l (I)

(9)

Since at the resonance the variable φ˜ has frequency 0, we have that ˜ = αk,l I 2 + O(I˜3 ) Hk,l (I) Furthermore, α will not be zero since it will be close to the second derivative of the unperturbed Hamiltonian, which we assumed is strictly positive (twist condition). Note that the dynamics of (9) is very similar to the dynamics of a pendulum with a potential of size ε . In this case, the variable A does not play any role at all. There will be homoclinic orbits to the maximum of the potential. These orbits will be given by the conservation of energy and the form of the kinetic energy as  (10) I˜ = ±ε 1/2 α −1 (maxV −V (φ˜ )) + O(ε ). Inside these curves, the system does a rotation. If the maximum is non-degenerate – another hypothesis which is easy to verify in practice and which holds for generic V – we see that the orbits described in (10) are orbits that start and end in a critical point, which is hyperbolic. They are at the same time the stable and the unstable manifolds of this hyperbolic fixed point. Note that these orbits are very different from the KAM tori. This is the reason why the KAM foliation gets interrupted by gaps of order ε 1/2 . It is important to remark that the stable and unstable manifolds of these periodic points have Lyapunov exponents O(ε 1/2 ). This is much smaller than the Lyapunov exponents in the transverse directions, which are independent of ε . Hence, when we talk about the stable manifolds restricted to Λ this is not the same as the W s in the sense of the theory of normally hyperbolic invariant manifolds , which requires convergence at an exponential rate of order 1. The dynamics of the averaged system – we will see that many of these features are preserved in the full system – consists of the foliation of – more or less horizontal – curves given by the orbits of the integrable system interrupted by a group of eyes or islands. At a resonance of type k, l we obtain k eyes. The amplitude of these eyes is O(ε 1/2 ). Remark 4.2. The above classification ignores some stripes of width O(ε ) near the separation of the regions. The conclusions remain valid if we realize that the separation between the zones – the choice of L – was a choice we made. We can repeat the same analysis with an slightly different L and see that the ambiguous zones are different in the two procedures. So that by doing the analysis twice with slightly different L one establishes the conclusions above for all the phase space. Remark 4.3. The choice of separation between the resonances zones is rather wasteful (even if it makes the estimates and the concepts easier). We assign the same

312

A. Delshams et al.

width to all the resonances even if it is clear that the real width will decrease with ε . (In particular, we expect that the optimal size would be close to ε 1/2 ). Furthermore, if the original Hamiltonian is several times differentiable, then, its Fourier coefficients will decrease at least like a power of k, l. Hence the Vk,l will become smaller with k, l. Hence, if for a fixed ε we decide to consider only resonant regions of size ε B , we only need to consider a finite number of resonances – which will grow as ε → 0 if B > 1/2. Considerations of these type were known heuristically since at least [Chi79]. A rigorous implementation appears in [DH06]. The paper [DH06] includes also considerations of repeated averaging – discussed in the next section – and a very detailed analysis of the motion in each resonance with error terms.

4.4 Repeated averaging The method of averaging can be applied several times. Indeed, in celestial mechanics it has been common for centuries to do at least two steps of averaging. In the region that was marked as integrable in the first step, after we perform the change of variables, we are left with a quasi-integrable system. The perturbation parameter is ε 2 . We can restart the procedure and get again some regions where the system can be made integrable up to O(ε 2 ) and new resonant regions in which the dynamics has eyes, which will now be of size ε rather than ε 1/2 . In the resonant regions, nothing much happens except that the resonant potential Vk,l gets deformed. In the case that the perturbation is a trigonometric polynomial, the number of resonances we get at each step is finite and given a number of steps, we can get an L which works for all cases. The result of applying averaging twice is depicted crudely in Figure 8.15 For future analysis, the only important thing is that near resonances, we encounter separatrices well approximated by other tori and that, outside the resonances the system is very approximately integrable.

4.5 Invariant objects generated by resonances: secondary tori, lower dimensional tori The resonant averaging described above, gives very accurate predictions of the dynamics. The difference between the perturbed system expressed in a system of coordinates and the true system – in a smooth norm – is smaller than CN ε N . The constants CN grow very fast. 15 We have ignored, for example, the fact that inside the big islands of size ε 1/2 there are other baby islands of size ε going around.

Geometric approaches to the problem of instability

313

Fig. 8 Schematic description of the predictions for the dynamics by the averaging method.

This can be taken advantage off in two different ways: A) If some perturbation theories apply, we can conclude that some of the invariant objects for the integrable system, persist for the true system. B) We have good control of some long orbits that, using some conditions can be glued together or shadowed. This can be applied to the two types of geometric programs mentioned in Section 1.1. In this section, we will be concerned mainly with point A) and will produce invariant objects. We will come to point B) in Section 6. If we consider the averaged system, we see that near resonances of order j, we obtain hyperbolic orbits, whose Lyapunov exponents are Cε j/2 + O(ε ( j+1)/2 ) and such that the angle between the stable and unstable directions are Cε j/2 + O(ε ( j+1)/2 ). Then, applying the implicit function theorem if N > j, we get that there are periodic orbits that persist.16 More importantly for our later applications we obtain that the stable and unstable manifolds are very similar to those of the integrable system. The results are depicted in Figure 9. We also can show that some of the quasi-periodic orbits with sufficiently large Diophantine constant persist. It is important to note that, one can get invariant tori of two types. One is tori which “go across”. These are the “primary tori” which are continuous deformations of the tori that were present in the unperturbed system. The tori inside the eyes of the resonance are of a completely different type. These are the “secondary tori” which were not present in the unperturbed system, but rather were created by the resonances. Note that as ε → 0, the eyes become flatter and the limit 16

There are many versions of this argument on persistence of periodic orbits. The basic idea goes back at least to Poincaré and Birkhoff.

314

A. Delshams et al.

e

e e3/2

e3/2

e3/2

e3/2

e1/2

e1/2 e3/2

Fig. 9 Illustration of an scaffolding of invariant objects in Λ . These invariant objects are ε 3/2 dense in the manifold.

of the tori is just a segment of periodic points. The tori merge with the stable and unstable manifolds. So that at the limit ε = 0 there is change of the topology. One point which is important is that there are invariant tori very close to the resonances both from the inside and from the outside. These problems had been considered in [Neı84, Her83] under slightly different hypothesis. The method used in [DLS06b] was, mainly, to study in detail the expansion of the action-angle coordinates in a neighborhood of the separatrix. Using the – more or less explicit – formulas one can find in textbooks, it is possible to show that the Cr norm of the change to action angle variables can be bounded by d −rA where A is an explicit number. As it turns out the twist constant does not degenerate – the frequency is singular, but in the good direction that the twist becomes infinite. On the other hand, remember that the error of the averaging method was less than CN ε N . It follows that one can apply the KAM theorem at a distance ε N/B . So that, one can get KAM tori – both rotating or librating – faster that a power of ε . The power is arbitrarily large assuming that the system is differentiable enough. The paper [DLS06b, Section 8] contained other considerations on properties of the KAM tori as graphs and how the set of KAM tori close to the invariant circle can be interpolated with the others where the averaging method is different. The problem is somewhat difficult because depending on how does one relate the ε to the distance to the separatrix, and to the fixed point, the expression of the KAM tori has different leading expressions. It seems possible that using more the geometric methods developed after [DLS06b] was written, many of these technical calculations can be eliminated or improved in many ways. A significant extension of the results can be found in [DH06]. Another line of argument that seems promising is the use of KAM theory without action angle coordinates – the singularity of the action angle variables

Geometric approaches to the problem of instability

315

and the different expressions in different regions is one of the source of problems – [dlLGJV05, FS07] so that one can prove directly the persistence of the orbits in the level sets of the averaged energy. We hope to come back to this. In summary, it is possible to show that one can get persistence of many of the orbits predicted by the averaging method. For our purposes, it is enough to claim that we get an scaffolding of orbits which are much closer that ε – the size of the effect of the scattering map.

4.6 Heteroclinic intersections between the invariant objects generated by resonances Now we want to argue that the objects discussed in the previous sections possess heteroclinic intersections. Since these objects have different topologies and very different characteristics, it is useful to use the scattering map and the argument discussed in Section 3.3. To establish this intersection, we just compute the image of these invariant objects under the scattering map and check whether one can verify (6). Given that we have computed rather explicitly the leading expansions of the scattering map and the leading expansions of the invariant objects, it is possible to compute the angles of intersections of manifolds. If these angles are not zero in the leading approximation, then, the implicit function theorem will establish that the true invariant manifolds satisfy (6). The effect of the scattering map on the invariant objects is depicted schematically in Figure 10.

e

e e3/2

e

e

e3/2

e3/2 e

e3/2

e3/2 e

e3/2

e e1/2

e e1/2 e3/2

e3/2 e3/2

Fig. 10 Effect of the scattering map on the invariant objects found in Figure 9.

e3/2

316

A. Delshams et al.

Therefore, the above calculation gives – rather explicit – expressions so that, if they do not vanish, then indeed we can obtain heteroclinic excursions between a primary torus below the resonance, to a secondary torus inside the resonance, and then to another torus above the resonance. The non-vanishing of these explicit expressions giving the angles is a nondegeneracy assumption on the perturbation. It is intuitively clear that the conditions hold rather generically. Basically, they are a comparison of two effects: the deformation of the invariant objects in Λ and the effect of the scattering map. We note that the first effect, is very much affected by the behavior of the perturbation near Λ , but not by the behavior of the perturbation near Γ . The scattering map has the opposite properties. Hence, if by some miracle, the angles happened to be zero, some perturbation near Γ could destroy this coincidence. Remark 4.4. The calculation of the scattering map in [DLS06b] was based on traditional methods of perturbations of slow variables. This had the consequence that only the energy component of the scattering map could be computed. The use of the symplectic properties, which was developed in [DLS06a] and explained in Section 3.7, simplifies and extends the calculation. Note also that we mentioned that the invariant objects are very close to the level sets of a function Ψε . Since the scattering map is a symplectic map close to the identity, the images of the level sets of Ψε will be level sets of the function Ψε + ε {Ψε , S0 } + h.o.t. See [DH06].

5 The method of correctly aligned windows The method of correctly aligned windows is a way of proving that given segments of orbits – with some extra conditions – one can get an orbit that tracks them. Since we never have to consider more than finite orbits, in principle, we do not need the existence of invariant objects. On the other hand, considerations about times become relevant. This is the reason why one gets explicit estimates on diffusion time. The method has its origins in [Eas78, EM79, Eas89]. The version we will discuss comes from [ZG04, GZ04]. One can think of a window, as a topological version of a rectangle with some marked sides. Windows are correctly aligned when the image of one stretches across the other. A window in a n-dimensional manifold M is a compact subset W of M together with a C0 -coordinate system (x, y) : U → Ru × Rs defined in neighborhood U of W , where u + s = n, such that the homeomorphic image of W through this coordinate system is the rectangle [0, 1]u ×[0, 1]s . The subset W − of W that corresponds through the coordinates (x, y) to ∂ [0, 1]u × [0, 1]s is called the ‘exit set’ and the subset W + of W that corresponds through the local coordinates (x, y) to [0, 1]u × ∂ [0, 1]s is called the ‘entry set’ of W . Here ∂ denotes the topological boundary of a set. If we want to specify the dimension u of the unstable-like direction and the dimension s of the

Geometric approaches to the problem of instability

317

stable-like direction of a window W , we refer to W as an (u, s)-window. We will assume that u > 0. Let W1 , W2 be two (u, s)-windows in M, and let (x1 , y1 ) : U1 → Rn and (x2 , y2 ) : U2 → Rn be the corresponding coordinates systems. Let f be a continuous map on M; we will denote its expression (x2 , y2 ) = f (x1 , y1 ) in local coordinates also by f . Assume f (U1 ) ⊆ U2 . We say that W1 is correctly aligned with W2 under f provided that the following conditions are satisfied: / (i) f (∂ [0, 1]u × [0, 1]s ) ∩ [0, 1]u × [0, 1]s = 0, ∂ [0, 1]s ) = 0. / (ii) there exists a point y0 ∈ [0, 1]s such that

f ([0, 1]u × [0, 1]s ) ∩ ([0, 1]u ×

(a) f ([0, 1]u × {y0 }) ⊆ int ([0, 1]u × [0, 1]s ∪ (Ru \ (0, 1)u ) × Rs ) , (b)

The map Ay0 : Ru → Ru defined by Ay0 (x) = π1 ( f (x, y0 )) satisfies Ay0 (∂ [0, 1]u ) ⊆ Ru \ [0, 1]u , deg(Ay0 , 0) = 0.

The main result is that “One can see through correctly aligned windows”. See [ZG04, GZ04]. Let Wi be a collection of (u, s)-windows in M, where i ∈ Z or i ∈ {0, . . . , d − 1}, with d > 0 (in the latter case, for convenience, we let Wi := W(i mod d) for all i ∈ Z). Let fi be a collection of continuous maps on M. If Wi is correctly aligned with Wi+1 , for all i, then there exists a point p ∈ W0 such that fi ◦ . . . ◦ f0 (p) ∈ Wi+1 , Moreover, if Wi+k = Wi for some k > 0 and all i, then the point p can be chosen so that fk−1 ◦ . . . ◦ f0 (p) = p.

318

A. Delshams et al.

If one takes very small windows, the behavior of the windows is determined by the derivative of the orbit. If the orbit is hyperbolic, by choosing the rectangles as products of balls along the stable direction and the unstable direction with the unstable being the exit direction, we can get the correct alignment. Then the result that one can see through chains of correctly aligned windows becomes the standard shadowing result. On the other hand, the method is more flexible since we can choose the sizes of the windows and the time we take to put them along the orbits. This has some advantages for non-uniformly hyperbolic systems. See the proof of the non-uniformly hyperbolic closing lemma in [Pol93]. On the other hand, the windows do not need to be small. As we will see in the next section, one can take advantage of large scale effects to get the alignment of windows. Notably, when one has some twist – shear – that causes some stretching, this can be used in place of the stretching caused by the hyperbolicity. It is also important to notice that, to check whether windows are well aligned or not, one can just study what happens on the boundary. In our applications the time of diffusion can be computed by the time that it takes the windows to stretch. An important technical tool [GL06a] is that, for systems that are close to product of systems, one can construct product windows and verify the alignment checking conditions on each of the factors.

6 The large gap model: the method of correctly aligned windows The method of correctly aligned windows has been applied to the large gap model in [GL06a, GL06b]. The construction of windows adapted to the problem of diffusion basically requires to choose the parameters of a sequence of windows (the length of the sides, the center in a good coordinate system) and choose the times taken to go from one to the next. Then, one has to verify that all the steps match. In practice this amounts to choosing two dozen of parameters and verifying about a dozen of trivial inequalities. Even if verifying the validity of the choices is not very hard, coming up with the good choices requires a good understanding of the behavior of the model. We now discuss some of the reasons behind the choices. We have already discussed the pseudo-orbits that appear. We go from the intersection to the manifold, rotate around and then escape back again. It is important to note that even the unperturbed system is not hyperbolic. The vector along the separatrices of the pendulum contracts both in the future and in the past. So that, these vectors in the intersection of the stable and unstable subspace and the forward Lyapunov exponent is different from the backward Lyapunov exponent.17 17 The equality of these two exponents was called regularity by Lyapunov and plays a very important role. See [BP01].

Geometric approaches to the problem of instability

319

The construction of windows, however, can take advantage of the fact that there are some direction with good hyperbolicity for a long stretch ( O(| ln(ε )|) ) of time while the orbit moves from Γ to Λ or back. The fact that one can control the behavior in the hyperbolic directions is possible because of the transversal intersection. (On the other hand, the windowing method, being a topological method could work with much weaker assumptions [GR04].) The treatment of the center directions is much more interesting. Of course, the windows that start close to Λ , go to Γ and come back to Λ are very well described by the scattering map. One does not have any hyperbolicity in these directions, but on the other hand, the twist does distort the windows and one can use this distortion to construct windows that are correctly aligned. This is very similar to the torsionhyperbolicity mechanism. In the paper, [GL06a] the windows were taken very thin in the action variables, but they were taken of order 1 in the angle. This allowed to avoid discussions of ergodization times and produced rather concrete estimates on the time. In [GL06b] the windows are chosen in a scale O(1/| ln(ε )|). This, of course, goes to zero, but it is much larger than the scales of the resonance. The orbits also do not come too close to the manifold. This has the effect that the method does not need to analyze what happens in the resonances. This method also leads to times of order O(ε | ln(ε )|) that – up to, perhaps, a constant – match the upper bounds obtained in [BBB03]. Similar results appear in [Tre04].

7 The large gap model in higher dimensions Some of the analysis in Section 4 can be adapted to higher dimensional models. See [DLS07]. We consider the same model as in (1), but now I, φ are higher dimensional variables. Again, for simplicity, for the moment, we assume that the perturbation h is a trigonometric polynomial. The averaging method described in Section 4.3 can be carried out pretty much the same way. The only difference is that now, that the resonances ω (I) · k = n are codimension 1 manifolds. If the number of degrees of freedom is more than 1, there will be intersections of these resonant surfaces. The intersection of two independent resonances are called multiple resonances. The multiplicity of the resonance – not to be confused with the order – is the dimension of the module of vectors k, n for which there is resonance relation. The order is the power of ε of the terms that cannot be eliminated. The mathematical analysis of multiple resonances and their role in diffusion remains a very interesting problem. Very important progress has been done in [Hal97, Hal99]. Nevertheless in [DLS07] it is argued that there exist diffusing orbits – under the assumption that h is polynomial – plus some non-degeneracy assumptions.

320

A. Delshams et al.

Fig. 11 Illustration of the paths of diffusion avoiding higher order resonances.

The key observation is that, under a twist condition, the multiple resonances can be contoured. (Since they happen on sets of codimension 2 or higher, there are paths that go around them.) The analysis of resonances of order 1 in higher dimensional systems is very similar to the analysis carried in Section 4.18 The upshot is that, under explicit non-degeneracy conditions, for any path in the space of actions that crosses only multiplicity one resonances, for 0 < ε small enough there orbits whose actions evolve along the path – up to errors that go to zero with ε .

8 Instability caused by normally hyperbolic laminations One of the standard heuristics in the numerical studies is that of modulational diffusion [Chi79, TLL80]. It is often described as saying that A degree of freedom becomes chaotic and drives another one. Mathematically, one can formulate this as perturbing a system which is the product of a system with some hyperbolic behavior, and another system which is integrable: Fε = Fh × Fi + O(ε ), where Fh (Λ ) = Λ and Λ is a hyperbolic set, and Fi : M → M is an integrable map. In the mathematical literature, some rigorous results have been obtained. The paper [MS02] constructed a specific system of this type. The paper [Moe02] used topological methods in two dimensions. Closely related to this paper is [EMR01]. 18

The scattering map does not require any change, but the persistence of tori of lower dimension becomes more complicated (one has to use KAM theory rather than the implicit function theorem). Also the secondary tori require some extra considerations.

Geometric approaches to the problem of instability

321

One systematic way to make sense of the above [Lla04, dlL06] is to observe that the set ∪x∈Λ {x} × M is a normally hyperbolic lamination for F0 . See Appendix A.3. The laminae are {x} × M are permuted by the map and the normal directions are hyperbolic. It was shown in [HPS77, Ch. 15] that these structures persist under perturbations in the sense that one can get slightly deformed collections of laminae which are also permuted under the map Fε . Of course, the dynamics on these laminae is not integrable anymore. The dynamics on the integrable parts is a random composition of maps, which one can consider as uncoupled as in [MS02].

8.1 Models with two time scales: geodesic flows, billiards with moving boundaries, Littlewood problems The above mechanism is particularly effective in systems that have two time scales. One important system is the model of a geodesic flow perturbed by a periodic or quasi-periodic potential considered by other methods in [Mat96, BT99, DLS00, DLS06c]. This dynamical system is defined on the cotangent bundle T ∗ M of a compact manifold M. It has the form: p˙ = −∇V (q, ω t),

q˙ = p,

(1)

where the potential V : M × Td and ω ∈ Rd is a non-resonant vector. When d = 1, the potential depends periodically on time. ˜ q= We note that the system satisfies some scaling properties. Setting p = ε p, q, ˜ t = ε −1t˜ and denoting by the derivative with respect to t˜, the system, above becomes ˜ εω t) q˜ = q˜ (2) p˜ = −ε 2 ∇V (q, So that, for high energy, the potential can be considered as a slow and weak perturbation. We will assume that the unperturbed geodesic flow has a horseshoe in the unit energy surface. Using the above scaling, we obtain that, considering the system for all the energies it possesses an invariant lamination. By the theory of persistence of normally hyperbolic invariant laminations, we obtain that this structure just gets deformed. If γ1 , . . . , γN are periodic orbits in the horseshoe, we denote |γi | the period and define:  1 |γi | ∂ V (γi (s),t) ds Gi (t) = |γi | 0 ∂ t This has the meaning of the gain of energy per unit time for orbits that stay in a  close proximity to the periodic orbit. Note that 01 Gi (t) dt = 0.

322

A. Delshams et al.

Fig. 12 Invariant normally hyperbolic laminations associated to the geodesic flow and the periodic geodesic flow.

Recall that, in the horseshoe, we have a symbolic dynamics for the hyperbolic orbits. That is, if we fix neighborhoods of these orbits, we can move from one to the other in arbitrary order. Each of the steps can be accomplished in a fixed time. By the persistence of the normally hyperbolic laminations, the same property persists when we consider the perturbation by the potential. So, we can switch from a neighborhood of an orbit to another one in a fixed time for the geodesic flow. For the potential, this is a slow time. In the periodic case, d = 1, we assume without loss of generality that V (q,t + 1) = V (q,t), ω = 1. If we assume that there exist 0 = a0 < a1 < · · · < aN = 1 in such a way that N

A≡∑

 ai

i=1 ai−1

Gi (t) dt > 0

(3)

then, we can construct orbits whose energy as function of time is larger than At − B. The idea is very simple. We stay close to γ1 during the macroscopic times [a0 , a1 ]. Using the symbolic dynamics, we can move to γ2 , etc. Hence, during a cycle, we have gained roughly A. In the quasi-periodic case, we just need to assume that it is possible to write to the rotation, Td = ∪Ni=1 Oi where Oi are sets with smooth boundary transversal  which only overlap in the boundary, and such that A ≡ ∑Ni=1 Oi Gi (τ ) d τ > 0. If we look at the symbolic dynamics, we see that the space of sequences that lead to linear gain in energy has positive Hausdorff dimension. Then, using that the conjugacy given by the stability, we obtain that, when the (3) are satisfied, the orbits with energy growing linearly are of positive Hausdorff dimension. It is shown in [Lla04] that if the metric is of negative curvature and, in case that it has dimension  3, that it satisfies some pinching conditions, then, the only C3

Geometric approaches to the problem of instability

323

Fig. 13 Illustration of the mechanisms of gain of energy based in locally hyperbolic manifolds.

potentials for which it is impossible to find orbits satisfying the hypothesis of the above result are the potentials of the form V (q,t) = V1 (q) +V2 (t). Very similar analysis applies to other systems which have two scales. One example is what we call the Littlewood models in higher dimensions. H(p, q,t) =

1 2 p +Vn (q) +Vm (q,t) 2

(4)

where p, q ∈ Rd , d  2, Vn , Vm are homogeneous of degree n, m respectively, n > m, n > 2,Vn > 0, Vm periodic or quasi-periodic in t. The fact that different terms have different homogeneities makes the geometric analysis similar to that of the geodesic flows. In the case d = 1, [Lit66a, Lit66b] constructed examples of potentials – which are not polynomials and with not very smooth dependence on time – with orbits with unbounded energy. Unfortunately, the papers contain a serious error. The papers [LL91, LZ95] showed that for terms which are like polynomials, and with smooth quasiperiodic perturbations the orbits stay bounded. An excellent survey of the history of these models and simplification of the results is [Lev92]. When the number of degrees of freedom is greater or equal than 2, a very similar analysis to the one carried out above for geodesic flows applies. We note that if we ˜ q = ε q, ˜ t = ε −m we get that the system (4) can be rewritten as: scale, p = ε m/2 p, 1 2 ˜ +Vm (q) H( p, ˜ q, ˜ t˜) = ( p) ˜ + ε 2m−nVn (q, ε mt) 2 so that the low degree polynomial can be considered as a small and slow perturbation and an analysis very similar to the one carried above for the geodesic flow applies. The only difference is that one gets that the orbits grow like a power. This is optimal due to a calculation in [LZ95].

324

A. Delshams et al.

One interesting example, which does not fit in the above theory proposed as a challenge by M. Levi is the system defined by a Hamiltonian 1 2 p + q61 + q41 + η q21 q22 + q1 f (t) 2 This is a challenging model because for large energy, the dominant term is the one degree of freedom system for which the theorem of [LZ95] applies. Another model which has scaling behavior is the billiard with moving boundaries. A higher dimensional model of the Fermi acceleration. For all these systems, when they are sufficiently chaotic, it seems possible to derive – heuristically – stochastic models for the growth of energy. These stochastic models can be analyzed rigorously and the final results compared satisfactorily with numerical simulations [DdlL06]. Even if parts of a stochastic theory of diffusion can be made rigorous, deriving a fully rigorous stochastic theory of diffusion remains a very challenging problem. Acknowledgements The work or M. G. and R. L has been supported by NSF grants. Visits of R. L to Barcelona were supported by ICREA. Some of the work reported here was supported ny MCyT-FEDER Grants BFM2003-9504 and MTM2006-00478. We thank the organizers of the Institute for a very warm hospitality (besides efficient organization). The participants in the institute made very valuable comments and suggestions. Particularly, R. Calleja, J. Mireles-James, P. Roldán and R. Treviño, helped in the preparation of the notes. H. Lomelí made many suggestions.

References [AA67] [AKN88] [Ale68a] [Ale68b] [Ale69] [Ale81] [AM78]

[Ang87] [AO98] [Arn63]

V.I. Arnold and A. Avez. Ergodic problems of classical mechanics. Benjamin, New York, 1967. V.I. Arnold, V.V. Kozlov, and A.I. Neishtadt. Dynamical Systems III, volume 3 of Encyclopaedia Math. Sci. Springer, Berlin, 1988. V. M. Alekseev. Quasirandom dynamical systems. I. Quasirandom diffeomorphisms. Mat. Sb. (N.S.), 76 (118):72–134, 1968. V. M. Alekseev. Quasirandom dynamical systems. II. One-dimensional nonlinear vibrations in a periodically perturbed field. Mat. Sb. (N.S.), 77 (119):545–601, 1968. V. M. Alekseev. Quasirandom dynamical systems. III. Quasirandom vibrations of one-dimensional oscillators. Mat. Sb. (N.S.), 78 (120):3–50, 1969. V.M. Alekseev. Quasirandom oscillations and qualitative questions in celestial mechanics. Transl., Ser. 2, Am. Math. Soc., 116:97–169, 1981. R. Abraham and J. E. Marsden. Foundations of mechanics. Benjamin/Cummings Publishing Co. Inc. Advanced Book Program, Reading, MA, 1978. Second edition, revised and enlarged, With the assistance of Tudor Ra¸tiu and Richard Cushman. Sigurd Angenent. The shadowing lemma for elliptic PDE. In Dynamics of infinitedimensional systems (Lisbon, 1986), pages 7–22. Springer, Berlin, 1987. J. Miguel Alonso and R. Ortega. Roots of unity and unbounded motions of an asymmetric oscillator. J. Differential Equations, 143(1):201–220, 1998. V. I. Arnol’d. Small denominators and problems of stability of motion in classical and celestial mechanics. Russian Math. Surveys, 18(6):85–191, 1963.

Geometric approaches to the problem of instability [Arn64]

325

V.I. Arnold. Instability of dynamical systems with several degrees of freedom. Sov. Math. Doklady, 5:581–585, 1964. [Arn89] V. I. Arnol d. Mathematical methods of classical mechanics. Springer, New York, second edition, 1989. Translated from the Russian by K. Vogtmann and A. Weinstein. [Ban88] V. Bangert. Mather sets for twist maps and geodesics on tori. In Dynamics reported, Vol. 1, pages 1–56. Teubner, Stuttgart, 1988. [BBB03] M. Berti, L. Biasco, and P. Bolle. Drift in phase space: a new variational mechanism with optimal diffusion time. J. Math. Pures Appl. (9), 82(6):613–664, 2003. [BCV01] U. Bessi, L. Chierchia, and E. Valdinoci. Upper bounds on Arnold diffusion times via Mather theory. J. Math. Pures Appl. (9), 80(1):105–129, 2001. [Bes96] U. Bessi. An approach to Arnol d’s diffusion through the calculus of variations. Nonlinear Anal., 26(6):1115–1135, 1996. [BK94] I. U. Bronstein and A. Ya. Kopanskiı. Smooth invariant manifolds and normal forms, volume 7 of World Scientific Series on Nonlinear Science. Series A: Monographs and Treatises. World Sci. Publ., River Edge, NJ, 1994. [BK05] J. Bourgain and V. Kaloshin. On diffusion in high-dimensional Hamiltonian systems. J. Funct. Anal., 229(1):1–61, 2005. [BLZ98] P. W. Bates, K. Lu, and C. Zeng. Existence and persistence of invariant manifolds for semiflows in Banach space. Mem. Amer. Math. Soc., 135(645):viii+129, 1998. [BP01] L. Barreira and Ya. Pesin. Lectures on Lyapunov exponents and smooth ergodic theory. In Smooth ergodic theory and its applications (Seattle, WA, 1999), pages 3– 106. Amer. Math. Soc., Providence, RI, 2001. Appendix A by M. Brin and Appendix B by D. Dolgopyat, H. Hu and Pesin. [BT79] M. Breitenecker and W. Thirring. Scattering theory in classical dynamics. Riv. Nuovo Cimento (3), 2(4):21, 1979. [BT99] S. Bolotin and D. Treschev. Unbounded growth of energy in nonautonomous Hamiltonian systems. Nonlinearity, 12(2):365–388, 1999. [Car81] J. R. Cary. Lie transform perturbation theory for Hamiltonian systems. Phys. Rep., 79(2):129–159, 1981. [CDMR06] E. Canalias, A. Delshams, J. Masdemont, and P. Roldán. The scattering map in the planar restricted three body problem. Celestial Mech. Dynam. Astronom., 95(1–4):155–171, 2006. [CG98] L. Chierchia and G. Gallavotti. Drift and diffusion in phase space. Ann. Inst. H. Poincaré Phys. Théor., 60(1):1–144, 1994. Erratum, Ann. Inst. H. Poincaré, Phys. Théor. 68 (1):135, 1998. [CG03] Jacky Cresson and Christophe Guillet. Periodic orbits and Arnold diffusion. Discrete Contin. Dyn. Syst., 9(2):451–470, 2003. [Chi79] B.V. Chirikov. A universal instability of many-dimensional oscillator systems. Phys. Rep., 52(5):264–379, 1979. [CI99] G. Contreras and R. Iturriaga. Global minimizers of autonomous Lagrangians. 22o Colóquio Brasileiro de Matemática. [22nd Brazilian Mathematics Colloquium]. Instituto de Matemática Pura e Aplicada (IMPA), Rio de Janeiro, 1999. [Con68] C. C. Conley. Low energy transit orbits in the restricted three-body problem. SIAM J. Appl. Math., 16:732–746, 1968. [Con78] C. Conley. Isolated invariant sets and the Morse index, volume 38 of CBMS Regional Conference Series in Mathematics. Amer. Math. Soc., Providence, RI, 1978. [Cre97] J. Cresson. A λ -lemma for partially hyperbolic tori and the obstruction property. Lett. Math. Phys., 42(4):363–377, 1997. [CY04a] C.-Q. Cheng and J. Yan. Arnold diffusion in Hamiltonian systems: 1 a priori unstable case. Preprint 04–265, [email protected], 2004. [CY04b] C.-Q. Cheng and J. Yan. Existence of diffusion orbits in a priori unstable Hamiltonian systems. J. Differential Geom., 67(3):457–517, 2004. [DdlL06] D. Dolgopyat and R. de la Llave. Stochastic acceleration. Manusrcript, 2006.

326

A. Delshams et al.

[DDLLS00] A. Delshams, R. de la Llave, and T. M. Seara. Unbounded growth of energy in periodic perturbations of geodesic flows of the torus. In Hamiltonian systems and celestial mechanics (Pátzcuaro, 1998), volume 6 of World Sci. Monogr. Ser. Math., pages 90–110. World Sci. Publ., River Edge, NJ, 2000. [DG96] A. Delshams and P. Gutiérrez. Effective stability and KAM theory. J. Differential Equations, 128(2):415–490, 1996. [DH06] A. Delshams and G. Huguet. The large gap problem in arnold diffusion for non polynomial perturbations of an a-priori unstable hamiltonian system. Manuscript, 2006. [DLC83] R. Douady and P. Le Calvez. Exemple de point fixe elliptique non topologiquement stable en dimension 4. C. R. Acad. Sci. Paris Sér. I Math., 296(21):895–898, 1983. [dlL06] R. de la Llave. Some recent progress in geometric methods in the instability problem in Hamiltonian mechanics. In International Congress of Mathematicians. Vol. II, pages 1705–1729. Eur. Math. Soc., Zürich, 2006. [dlLGJV05] R. de la Llave, A. González, À. Jorba, and J. Villanueva. KAM theory without action-angle variables. Nonlinearity, 18(2):855–895, 2005. [dlLRR07] R. de la Llave and R. Ramirez-Ros. Instability in billiards with moving boundaries. Manuscript, 2007. [DLS00] A. Delshams, R. de la Llave, and T.M. Seara. A geometric approach to the existence of orbits with unbounded energy in generic periodic perturbations by a potential of generic geodesic flows of T2 . Comm. Math. Phys., 209(2):353–392, 2000. [DLS03] Amadeu Delshams, Rafael de la Llave, and Tere M. Seara. A geometric mechanism for diffusion in Hamiltonian systems overcoming the large gap problem: announcement of results. Electron. Res. Announc. Amer. Math. Soc., 9:125–134 (electronic), 2003. [DLS06a] A. Delshams, R. de la Llave, and T. M. Seara. Geometric properties of the scattering map to a normally hyperbolic manifold. Adv. Math., 2006. To appear. [DLS06b] A. Delshams, R. de la Llave, and T. M. Seara. A geometric mechanism for diffusion in Hamiltonian systems overcoming the large gap problem: heuristics and rigorous verification on a model. Mem. Amer. Math. Soc., 179(844):viii+141, 2006. [DLS06c] A. Delshams, R. de la Llave, and T. M. Seara. Orbits of unbounded energy in quasiperiodic perturbations of geodesic flows. Adv. Math., 202(1):64–188, 2006. [DLS07] A. Delshams, R. de la Llave, and T. M. Seara. Instability of high dimensional hamiltonian systems: multiple resonances do not impede diffusion. 2007. [Dou88] R. Douady. Regular dependence of invariant curves and Aubry-Mather sets of twist maps of an annulus. Ergodic Theory Dynam. Systems, 8(4):555–584, 1988. [Dou89] R. Douady. Systèmes dynamiques non autonomes: démonstration d’un théorème de Pustyl’nikov. J. Math. Pures Appl. (9), 68(3):297–317, 1989. [DR97] A. Delshams and R. Ramírez-Ros. Melnikov potential for exact symplectic maps. Comm. Math. Phys., 190:213–245, 1997. [Eas78] R. W. Easton. Homoclinic phenomena in Hamiltonian systems with several degrees of freedom. J. Differential Equations, 29(2):241–252, 1978. [Eas89] R. W. Easton. Isolating blocks and epsilon chains for maps. Phys. D, 39(1):95–110, 1989. [EM79] R. W. Easton and R. McGehee. Homoclinic phenomena for orbits doubly asymptotic to an invariant three-sphere. Indiana Univ. Math. J., 28(2):211–240, 1979. [EMR01] R. W. Easton, J. D. Meiss, and G. Roberts. Drift by coupling to an anti-integrable limit. Phys. D, 156(3-4):201–218, 2001. [Fen72] N. Fenichel. Persistence and smoothness of invariant manifolds for flows. Indiana Univ. Math. J., 21:193–226, 1971/1972. [Fen77] N. Fenichel. Asymptotic stability with rate conditions. II. Indiana Univ. Math. J., 26(1):81–93, 1977. [Fen79] N. Fenichel. Geometric singular perturbation theory for ordinary differential equations. J. Differential Equations, 31(1):53–98, 1979.

Geometric approaches to the problem of instability [Fen74] [FGL05]

[FLG06]

[FM00] [FM01] [FM03] [FS07] [Gar00] [GL06a] [GL06b] [GLF05]

[GR04]

[GZ04] [Hal97] [Hal99] [HdlL06a] [HdlL06b]

[HdlL06c]

[HdlL07]

[Hed32] [Her83] [HM82] [HP70]

327

N. Fenichel. Asymptotic stability with rate conditions. Indiana Univ. Math. J., 23:1109–1137, 1973/74. C. Froeschlé, M. Guzzo, and E. Lega. Local and global diffusion along resonant lines in discrete quasi-integrable dynamical systems. Celestial Mech. Dynam. Astronom., 92(1–3):243–255, 2005. C. Froeschlé, E. Lega, and M. Guzzo. Analysis of the chaotic behaviour of orbits diffusing along the Arnold web. Celestial Mech. Dynam. Astronom., 95(1–4):141–153, 2006. E. Fontich and P. Martín. Differentiable invariant manifolds for partially hyperbolic tori and a lambda lemma. Nonlinearity, 13(5):1561–1593, 2000. E. Fontich and P. Martín. Arnold diffusion in perturbations of analytic integrable Hamiltonian systems. Discrete Contin. Dynam. Systems, 7(1):61–84, 2001. E. Fontich and P. Martín. Hamiltonian systems with orbits covering densely submanifolds of small codimension. Nonlinear Anal., 52(1):315–327, 2003. R. Fontich, E. de la Llave and Y. Sire. Construction of invariant whiskered tori by a parameterization method. 2007. Manuscript. Antonio García. Transition tori near an elliptic fixed point. Discrete Contin. Dynam. Systems, 6(2):381–392, 2000. M. Gidea and R. de la Llave. Arnold diffusion with optimal time in the large gap problem. Preprint, 2006. M. Gidea and R. de la Llave. Topological methods in the instability problem of Hamiltonian systems. Discrete Contin. Dyn. Syst., 14(2):295–328, 2006. M. Guzzo, E. Lega, and C. Froeschlé. First numerical evidence of global Arnold diffusion in quasi-integrable systems. Discrete Contin. Dyn. Syst. Ser. B, 5(3):687–698, 2005. M. Gidea and C. Robinson. Symbolic dynamics for transition tori II. In New advances in celestial mechanics and Hamiltonian systems, pages 95–109. Kluwer, Dordrecht, The Netherlands, 2004. M. Gidea and P. Zgliczy´nski. Covering relations for multidimensional dynamical systems. II. J. Differential Equations, 202(1):59–80, 2004. G. Haller. Universal homoclinic bifurcations and chaos near double resonances. J. Statist. Phys., 86(5-6):1011–1051, 1997. G. Haller. Chaos near resonance. Springer, New York, 1999. À. Haro and R. de la Llave. Manifolds on the verge of a hyperbolicity breakdown. Chaos, 16(1):013120, 8, 2006. À. Haro and R. de la Llave. A parameterization method for the computation of invariant tori and their whiskers in quasi-periodic maps: numerical algorithms. Discrete Contin. Dyn. Syst. Ser. B, 6(6):1261–1300 (electronic), 2006. A. Haro and R. de la Llave. A parameterization method for the computation of invariant tori and their whiskers in quasi-periodic maps: rigorous results. J. Differential Equations, 228(2):530–579, 2006. A. Haro and R. de la Llave. A parameterization method for the computation of invariant tori and their whiskers in quasi-periodic maps: explorations and mechanisms for the breakdown of hyperbolicity. SIAM J. Appl. Dyn. Syst., 6(1):142–207 (electronic), 2007. G. A. Hedlund. Geodesics on a two-dimensional Riemannian manifold with periodic coefficients. Ann. of Math., 33:719–739, 1932. M. R. Herman. Sur les courbes invariantes par les difféomorphismes de l’anneau. Vol. 1, volume 103 of Astérisque. Société Mathématique de France, Paris, 1983. P. J. Holmes and J. E. Marsden. Melnikov’s method and Arnold diffusion for perturbations of integrable Hamiltonian systems. J. Math. Phys., 23(4):669–675, 1982. M. W. Hirsch and C. C. Pugh. Stable manifolds and hyperbolic sets. In S. Chern and S. Smale, editors, Global Analysis (Proc. Sympos. Pure Math., Vol. XIV, Berkeley, Calif., 1968), pages 133–163, Amer. Math. Soc. Providence, RI, 1970.

328 [HPS77] [Hun68] [HW99] [KP90]

[KPT95] [Lev92]

[Lit66a]

[Lit66b] [LL91] [Lla01]

[Lla02] [Lla04] [LM88] [LMM86]

[LMS03]

[LP66] [LT83]

[LY97] [LZ95] [Mañ97] [Mat93] [Mat96] [Mey91]

A. Delshams et al. M. W. Hirsch, C. C. Pugh, and M. Shub. Invariant manifolds, volume 583 of Lecture Notes in Math. Springer, Berlin, 1977. W. Hunziker. The s-matrix in classical mechanics. Com. Math. Phys., 8(4):282–299, 1968. B. Hasselblatt and A. Wilkinson. Prevalence of non-Lipschitz Anosov foliations. Ergodic Theory Dynam. Systems, 19(3):643–656, 1999. U. Kirchgraber and K. J. Palmer. Geometry in the neighborhood of invariant manifolds of maps and flows and linearization. Longman Scientific & Technical, Harlow, 1990. T. Krüger, L. D. Pustyl nikov, and S. E. Troubetzkoy. Acceleration of bouncing balls in external fields. Nonlinearity, 8(3):397–410, 1995. M. Levi. On Littlewood’s counterexample of unbounded motions in superquadratic potentials. In Dynamics reported: expositions in dynamical systems, pages 113–124. Springer, Berlin, 1992. J. E. Littlewood. Unbounded solutions of an equation y¨ + g(y) = p(t), with p(t) periodic and bounded, and g(y)/y → ∞ as y → ±∞. J. London Math. Soc., 41:497– 507, 1966. J. E. Littlewood. Unbounded solutions of y¨ + g(y) = p(t). J. London Math. Soc., 41:491–496, 1966. S. Laederich and M. Levi. Invariant curves and time-dependent potentials. Ergodic Theory Dynam. Systems, 11(2):365–378, 1991. R. de la Llave. A tutorial on KAM theory. In Smooth ergodic theory and its applications (Seattle, WA, 1999), pages 175–292. Amer. Math. Soc., Providence, RI, 2001. R. de la Llave. Orbits of unbounded energy in perturbations of geodesic flows by periodic potentials. a simple construction. Preprint, 2002. R. de la Llave. Orbits of unbounded energy in perturbation of geodesic flows: a simple mechanism. Preprint, 2004. P. Lochak and C. Meunier. Multiphase Averaging for Classical Systems, volume 72 of Appl. Math. Sci. Springer, New York, 1988. R. de la Llave, J. M. Marco, and R. Moriyón. Canonical perturbation theory of Anosov systems and regularity results for the Livšic cohomology equation. Ann. of Math. (2), 123(3):537–611, 1986. P. Lochak, J.-P. Marco, and D. Sauzin. On the splitting of invariant manifolds in multidimensional near-integrable Hamiltonian systems. Mem. Amer. Math. Soc., 163(775):viii+145, 2003. M. de La Place. Celestial mechanics. Vols. I–IV. Translated from the French, with a commentary, by Nathaniel Bowditch. Chelsea Publishing, Bronx, NY, 1966. M. A. Lieberman and Jeffrey L. Tennyson. Chaotic motion along resonance layers in near-integrable Hamiltonian systems with three or more degrees of freedom. In C. Wendell Horton, Jr. and L. E. Reichl, editors, Long-time prediction in dynamics (Lakeway, Tex., 1981), pages 179–211. Wiley, New York, 1983. M. Levi and J. You. Oscillatory escape in a Duffing equation with a polynomial potential. J. Differential Equations, 140(2):415–426, 1997. M. Levi and E. Zehnder. Boundedness of solutions for quasiperiodic potentials. SIAM J. Math. Anal., 26(5):1233–1256, 1995. R. Mañé. Lagrangian flows: the dynamics of globally minimizing orbits. Bol. Soc. Brasil. Mat. (N.S.), 28(2):141–153, 1997. J. N. Mather. Variational construction of connecting orbits. Ann. Inst. Fourier (Grenoble), 43(5):1349–1386, 1993. J. N. Mather. Graduate course at Princeton, 95–96, and Lectures at Penn State, Spring 96, Paris, Summer 96, Austin, Fall 96. K. R. Meyer. Lie transform tutorial. II. In Kenneth R. Meyer and Dieter S. Schmidt, editors, Computer aided proofs in analysis (Cincinnati, OH, 1989), volume 28 of IMA Vol. Math. Appl., pages 190–210. Springer, New York, 1991.

Geometric approaches to the problem of instability [Moe96] [Moe02]

[Moe05] [Mor24] [Mos69] [Mos73]

[MS02]

[Neı84] [Nel69] [Nie07] [Ort97] [Ort04] [Pal00] [Pes04] [Pil99] [Poi99] [Pol93]

[PS70] [Pus77a] [Pus77b]

[Pus95]

[Rob71] [Rob88] [Rob02]

329

R. Moeckel. Transition tori in the five-body problem. J. Differential Equations, 129(2):290–314, 1996. R. Moeckel. Generic drift on Cantor sets of annuli. In Celestial mechanics (Evanston, IL, 1999), volume 292 of Contemp. Math., pages 163–171. Amer. Math. Soc., Providence, RI, 2002. R. Moeckel. A variational proof of existence of transit orbits in the restricted threebody problem. Dyn. Syst., 20(1):45–58, 2005. M. Morse. A fundamental class of geodesics on any closed surface of genus greater than one. Trans. Amer. Math. Soc., 26:26–60, 1924. J. Moser. On a theorem of Anosov. J. Differential Equations, 5:411–440, 1969. J. Moser. Stable and random motions in dynamical systems. Princeton University Press, Princeton, NJ, 1973. With special emphasis on celestial mechanics, Hermann Weyl Lectures, the Institute for Advanced Study, Princeton, NJ, Annals Math. Studies, No. 77. J.-P. Marco and D. Sauzin. Stability and instability for Gevrey quasi-convex nearintegrable Hamiltonian systems. Publ. Math. Inst. Hautes Études Sci., (96):199–275 (2003), 2002. A. I. Neıshtadt. The separation of motions in systems with rapidly rotating phase. J. Appl. Math. Mech., 48(2):133–139, 1984. E. Nelson. Topics in dynamics. I: Flows. Mathematical Notes. Princeton University Press, Princeton, NJ, 1969. L. Niederman. Prevalence of exponential stability among nearly integrable Hamiltonian systems. Ergodic Theory Dynam. Systems, 27(3):905–928, 2007. R. Ortega. Nonexistence of invariant curves of mappings with small twist. Nonlinearity, 10(1):195–197, 1997. R. Ortega. Unbounded motions in forced newtonian equations. Preprint, 2004. K. Palmer. Shadowing in dynamical systems, volume 501 of Mathematics and its Applications. Kluwer, Dordrecht, The Netherlands, 2000. Theory and applications. Y. B. Pesin. Lectures on partial hyperbolicity and stable ergodicity. Zurich Lectures in Advanced Mathematics. Eur. Math. Soc. (EMS), Zürich, 2004. S. Yu. Pilyugin. Shadowing in dynamical systems, volume 1706 of Lecture Notes in Mathematics. Springer, Berlin, 1999. H. Poincaré. Les méthodes nouvelles de la mécanique céleste, volume 1, 2, 3. Gauthier-Villars, Paris, 1892–1899. M. Pollicott. Lectures on ergodic theory and Pesin theory on compact manifolds, volume 180 of London Mathematical Society Lecture Note Series. Cambridge University Press, Cambridge, 1993. C. Pugh and M. Shub. Linearization of normally hyperbolic diffeomorphisms and flows. Invent. Math., 10:187–198, 1970. L. D. Pustyl’nikov. Stable and oscillating motions in nonautonomous dynamical systems. II. Trudy Moskov. Mat. Obšˇc., 34:3–103, 1977. L. D. Pustyl’nikov. Stable and oscillating motions in nonautonomous dynamical systems. II. Trudy Moskov. Mat. Obšˇc., 34:3–103, 1977. English translation: Trans. Moscow Math. Soc., 1978, Issue 2, pages 1–101. Amer. Math. Soc., Providence, RI, 1978, L. D. Pustyl’nikov. Poincaré models, rigorous justification of the second law of thermodynamics from mechanics, and the Fermi acceleration mechanism. Uspekhi Mat. Nauk, 50(1(301)):143–186, 1995. C. Robinson. Differentiable conjugacy near compact invariant manifolds. Bol. Soc. Brasil. Mat., 2(1):33–44, 1971. C. Robinson. Horseshoes for autonomous Hamiltonian systems using the Melnikov integral. Ergodic Theory Dynam. Systems, 8:395–409, 1988. C. Robinson. Symbolic dynamics for transition tori. In Celestial mechanics (Evanston, IL, 1999), volume 292 of Contemp. Math., pages 199–208. Amer. Math. Soc., Providence, RI, 2002.

330 [RS02]

[Sac65] [Shu78] [Sim99] [Sit53] [Ten82] [Thi83]

[TLL80]

[Tre02a] [Tre02b] [Tre04] [Wei73] [Wei79]

[Zas02] [ZG04] [ZZN+ 89]

A. Delshams et al. P. H. Rabinowitz and E. W. Stredulinsky. A variational shadowing method. In Celestial mechanics (Evanston, IL, 1999), volume 292 of Contemp. Math., pages 185–197. Amer. Math. Soc., Providence, RI, 2002. R. J. Sacker. A new approach to the perturbation theory of invariant surfaces. Comm. Pure Appl. Math., 18:717–732, 1965. M. Shub. Stabilité globale des systèmes dynamiques. Société Mathématique de France, Paris, 1978. With an English preface and summary. C. Simó, editor. Hamiltonian systems with three or more degrees of freedom, Kluwer, Dordrecht, The Netherlands, 1999. K. A. Sitnikov. On the possibility of capture in the problem of three bodies. Mat. Sbornik N.S., 32(74):693–705, 1953. J. Tennyson. Resonance transport in near-integrable systems with many degrees of freedom. Phys. D, 5(1):123–135, 1982. W. Thirring. Classical scattering theory. In Conference on differential geometric methods in theoretical physics (Trieste, 1981), pages 41–64. World Sci. Publ., Singapore, 1983. J. L. Tennyson, M. A. Lieberman, and A. J. Lichtenberg. Diffusion in near-integrable Hamiltonian systems with three degrees of freedom. In Melvin Month and John C. Herrera, editors, Nonlinear dynamics and the beam-beam interaction (Sympos., Brookhaven Nat. Lab., New York, 1979), pages 272–301. Amer. Inst. Physics, New York, 1980. D. Treschev. Multidimensional symplectic separatrix maps. J. Nonlinear Sci., 12(1):27–58, 2002. D. Treschev. Trajectories in a neighbourhood of asymptotic surfaces of a priori unstable Hamiltonian systems. Nonlinearity, 15(6):2033–2052, 2002. D. Treschev. Evolution of slow variables in a priori unstable Hamiltonian systems. Nonlinearity, 17(5):1803–1841, 2004. A. Weinstein. Lagrangian submanifolds and Hamiltonian systems. Ann. of Math. (2), 98:377–410, 1973. A. Weinstein. Lectures on symplectic manifolds, volume 29 of CBMS Regional Conference Series in Mathematics. Amer. Math. Soc. Providence, RI, 1979. Corrected reprint. G. M. Zaslavsky. Chaos, fractional kinetics, and anomalous transport. Phys. Rep., 371(6):461–580, 2002. P. Zgliczy´nski and M. Gidea. Covering relations for multidimensional dynamical systems. J. Differential Equations, 202(1):32–58, 2004. G. M. Zaslavskiı, M. Yu. Zakharov, A. I. Neıshtadt, R. Z. Sagdeev, D. A. Usikov, and A. A. Chernikov. Multidimensional Hamiltonian chaos. Zh. Èksper. Teoret. Fiz., 96(11):1563–1586, 1989.

Appendix A: Normally hyperbolic manifolds In this section, we recall some results in the literature on normally hyperbolic manifolds. Good references are [Fen72, Fen74, Fen77, HPS77, Pes04]. For simplicity, we will discuss only the case of diffeomorphisms. The case of flows is very similar. For many of the applications (persistence of invariant manifolds, regularity) the case of flows follows from the case of diffeomorphism by taking time-1 maps.

Geometric approaches to the problem of instability

331

Let M be a smooth d-dimensional manifold, f : M → M a Cr diffeomorphism, r  1. Definition 9.1. Let Λ ⊂ M be a C1 submanifold invariant under f , f (Λ ) = Λ . We say that Λ is a normally hyperbolic invariant manifold if there exist a constant C > 0, rates 0 < λ < µ −1 < 1 and a splitting for every x ∈ Λ Tx M = Exs ⊕ Exu ⊕ TxΛ in such a way that v ∈ Exs ⇔ |D f n (x)v|
Cλ n |v| v ∈ Exu

|n|

⇔ |D f (x)v|
Cλ |v| n

|n|

v ∈ TxΛ ⇔ |D f (x)v|
C µ |v| n

n0 n
0

(1)

n∈Z

In this exposition, we will assume that Λ is compact and, without loss of generality, connected. Remark 9.1. The set up can be weakened in several directions which appear in applications. For example, as remarked in [HPS77], instead of assuming that Λ is compact, it suffices to assume that f is Cr in a neighborhood of Λ with all the derivatives of order up to r uniformly bounded. The non-compact case involves some complications such as study of extension operators. These considerations become much more important in the extension of the theory to infinite dimensional Banach spaces, which we will also not consider [BLZ98]. In these infinite dimensional cases, the standard arguments often give one or two derivatives less in the conclusions than the finite dimensional compact arguments. We also note that some parts of the theory are also true for manifolds with boundary such that f (Λ ) ⊂ Λ , d( f (∂Λ ), ∂Λ ) > 0 (inflowing) or f (Λ ) ⊂ Λ , d( f (∂Λ ), ∂Λ ) > 0 (outflowing). Note that the definition of stable (resp. unstable) directions in (1) requires serious changes in the outflowing (resp. inflowing) cases. An adaptation of the theory to the inflowing and outflowing cases is done in [Fen72]. Note that, even if these definitions become possible, the resulting objects may lack some of the properties of the more standard definitions. For example, the stable spaces are not unique in the inflowing case, so that issues of regularity are more delicate, even if well understood in the literature. In some applications to instability, one often gets systems with two time scales, so that the hyperbolicity degenerates. Therefore it is useful to keep explicit track of how C, λ , µ , the parameters affecting the quality of the hyperbolicity in (1) enter in the hypothesis of the theorems. See [Fen79]. A self-contained detailed treatment of a case that involves several of these complications can be found in Appendix A of [DLS06c]. It follows from (1) that Exs , Exu depend continuously on x. In particular, the dimension of Exs , Exu are independent of x. In fact, using the invariant section theorem [HP70] or some direct arguments [Fen74, Fen77] they are C−1 ,

332

A. Delshams et al.

  | log λ | .  < min r, log µ

(2)

Indeed, using some variants of these arguments, it is possible to show that the invariant manifold Λ is C – even if the hypothesis of the definition only require it is C1 . In general, one cannot improve on these regularities. [Mos69] contains explicit examples – even trigonometric polynomials – where the regularity claimed above is sharp, and [HW99] shows that this regularity is indeed sharp for generic examples. Hence, in general, one cannot expect that the normally hyperbolic invariant manifolds are C∞ even if f is a polynomial. One can however have uniform lower bounds for all the Cr maps which are in a C1 neighborhood. The regularity of overflowing (resp. inflowing) manifolds is even more problematic since the stable (resp. unstable) bundles are not uniquely defined, hence the hyperbolicity constants do not have a unique value. Given a normally hyperbolic invariant manifold Λ we define WΛs = {y ∈ M | d( f n (y), Λ )
Cy λ n , n  0} WΛu = {y ∈ M | d( f n (y), Λ )
Cy λ |n| , n
0} Furthermore, for each x ∈ Λ , we define Wxs = {y ∈ M | d( f n (x), f n (y))
Cx,y λ n , n  0} Wxu = {y ∈ M | d( f n (x), f n (y))
Cx,y λ |n| , n
0} and we note that Exs = TxWxs and Exu = TxWxu . It is a fact that WΛs = WΛu =

/

Wxs

x∈Λ

/

Wxu

(3)

x∈Λ

Moreover, x = x˜ ⇒ Wxs ∩Wx˜s = 0, / Wxu ∩Wx˜u = 0. / The decomposition (3) can expressed geometrically saying that {Wxs }x∈Λ , {Wxu }x∈Λ are a foliation of WΛs , WΛu , respectively. We will refer to these foliations as Fs , Fu . Dynamically, the above statement means that, when the orbit of a point is approaching Λ , it approaches the orbit of a single point. This, as well as the uniqueness can be established easily by noting that, for two points in Λ , we have d( f n (x), f n (x))  C µ −n . Since λ µ < 1, we can see that if we fix y there can only be one x such that d( f n (x), f n (y))
Cλ n . We recall that in these circumstances we have that 1. Λ is a C manifold with  given in (2). 2. WΛs , WΛu are C−1 manifolds 3. Wxs , Wxu are Cr manifolds

Geometric approaches to the problem of instability

333

4. The maps x → Wxs , Wxu are C−1− j , when Wxs , Wxu are given the C j topologies in compact sets. 5. When x ∈ Λ , we have TxWΛs,u = Exs,u

TxWxs,u = Exs,u

6. As a consequence of the above, using the implicit function theorem, we have: Denote by WΛs,δ a δ -neighborhood of Λ in WΛs and by Wxs,δ a δ neighborhood of x in Wxs . Then, for sufficiently small δ , there is a C−1 diffeomorphism hs from WΛs,δ to a neighborhood of the zero section in E s . Furthermore, hs (Wxs ) ⊂ Exs . Note, that, even if Wxs are as smooth as the map, the dependence of the point on the base point has only some finite regularity that depends on the regularity exponents entering in (1). The manifold WΛs is invariant. That is f (WΛs ) = WΛs . Analogously, of course, the unstable manifolds. On the other hand, the manifolds Wxs are not invariant. They, however satisfy a covariance property (4) f (Wxs ) = W fs(x) The local behavior in a neighborhood of a normally hyperbolic invariant manifold is described very precisely by the following theorem in [HPS77, PS70], who show who show that if Λ is a normally hyperbolic invariant manifold , then there is a homeomorphism h from a neighborhood of the zero section in T Λ to a neighborhood in Λ in such a way that if x ∈ Λ , η ∈ Tx M and |η | is sufficiently small, we have (5) f ◦ h(x, η ) = h( f (x), D f (x)η ) The homeomorphism h is, of course, highly non-unique. Note that, in the case that Λ is just a point, the theorem reduces to the celebrated Hartman–Grobman theorem. Indeed the proof of the references above, after some clever reductions, becomes the Hartman–Grobman theorem in infinite dimensions. An important consequence of the linearization theorem is that if Λ is a normally hyperbolic invariant manifold , then, for any sufficiently small open neighborhood U of Λ we have / Λ= f n (U) '

n∈Z

Of course, if Λ ⊂ V ⊂ U, then Λ = n∈Z f n (V ). The homeomorphism h solving (5) is not unique and there are really terrible choices.19 Nevertheless, there are choices which are continuous and indeed Hölder in some of the variables. We also have that, Wxs,uloc = h(x, Exs,u ∪ Bδ ). The linearization (5) is a generalization of Hartman–Grobman theorem. Under appropriate non-resonance conditions on the possible rates of growth of the vectors 19 The lovers of pathologies can amuse themselves using the axiom of choice – Argh!! – to produce h solving (5) which are not measurable.

334

A. Delshams et al.

on Tx M|x∈Λ it is possible to obtain more precise linearizations [Rob71,KP90,BK94]. In contrast with the Sternberg Linearization theorem, the non-resonance conditions can fail in C1 open sets of diffeomorphisms. When the conditions for the linearization apply, then one can obtain very good estimates for the orbits that “fly by” the invariant manifold. In particular, one can get very detailed information about the separatrix map. Note that the time that one can spend in a “fly by” is unbounded, so that linearization gives information over trajectories that go over a long time.

A.1 Persistence and dependence on parameters One of the most important results of the theory of normally hyperbolic invariant manifolds is that they persist under perturbations and that they depend smoothly under parameters. Persistence means, roughly, that if a map f has an invariant manifold Λ f and g is sufficiently C1 close to f , then g also has an invariant manifold Λg . In these cases, the results on dependence on parameters and can be obtained very economically from the results on persistence by considering an extended system. Let f (x, ε ) : M × Σ → M is a family of maps (ε is the parameter). 0 such that

Variational methods for the problem of Arnold diffusion

355

T˜i0 ,T˜i1 ∞ min h∞ ci (ξ , m0 ) + hηi ,µi ,ψi ,e1 (m0 , m1 ) + hci+1 (m1 , ζ ) :

+ ) (m0 , m1 ) ∈ ∂ (Vi− ×Vi+1 T˜i0 ,T˜i1 ∞  min h∞ ci (ξ , m0 ) + hηi ,µi ,ψi ,e1 (m0 , m1 ) + hci+1 (m1 , ζ ) :

+ + 5εi∗ (m0 , m1 ) ∈ Vi− ×Vi+1

(2)

where ξ ∈ M0 (ci ), ζ ∈ M (c ). ' 0 i+1 For each integer i ∈ i j ∈Λ2 {i j , i j + 1, · · · , i j+1 − 1}, the situation is similar, the ˜ difference is that we do not need to lift M to its covering M. 1. There exists a local minimal orbit of the second type d γi : R → T M such that it solves the the Euler–Lagrange equation determined by L, α (d γi ) ⊂ A˜(ci ) and ω (d γi ) ⊂ A˜(ci+1 ): 2. Given a small number λi there is a non-negative function ψi (x,t)
λi such that ψi = 0 when t ∈ (−∞, 0] ∪ [1, ∞). For each fixed t, the support of ψi is contained in a small neighborhood of the open disk Oi and ψi = constant when it is restricted in Oi . Oi ∩ (N (ci )|0
t
t0 \(A (ci ) + δ )) = ∅,

∂ Oi ∩ N (ci )|0
t
t0 = ∅,

Oi ∩ (A (ci ) + δ ) = ∅;

3. There exist a closed 1-forms ηi with [ηi ] = ci and a U step 1-form µi such that the restriction on {t  t0 } is a closed 1-form µ¯ i on M with [µ¯ i ] = ci+1 − ci . The support of µi is disjoint with Oi . According to lemma 15, we can see that the set C˜ηi ,µi ,ψi (M) has the property: ∅ = Cηi ,µi ,ψi (M) ⊂ Oi ,

(3)

each orbit d γ (t) ∈ C˜ηi ,µi ,ψi (M) determines a local minimal orbit of L of the second type, which connects A˜(ci ) to A˜(ci+1 ). There are A˜cij ⊂ A˜(ci ), A˜cki+1 ⊂ A˜(ci+1 ) and an orbit d γ ∈ C˜η ,µ ,ψ (M) such that α (d γ ) ⊂ A˜cj and ω (d γ ) ⊂ A˜ck . Consei

i

i

i

i+1

+ quently, there exist two open disks Vi− and Vi+1 with V¯i− ⊂ (Acij |t=0 + δ )\A0 (ci ), + ⊂ (Acki+1 |t=0 + δ )\A0 (ci+1 ), two positive integers T˜i0 , T˜i1 and a positive small V¯i+1 number εi∗ > 0 such that

T˜i0 ,T˜i1 ∞ min h∞ ci (ξ , m0 ) + hηi ,µi ,ψi (m0 , m1 ) + hci+1 (m1 , ζ ) :

+ ) (m0 , m1 ) ∈ ∂ (Vi− ×Vi+1 T˜i0 ,T˜i1 ∞  min h∞ ci (ξ , m0 ) + hηi ,µi ,ψi (m0 , m1 ) + hci+1 (m1 , ζ ) :

+ + 5εi∗ (m0 , m1 ) ∈ Vi− ×Vi+1 where ξ ∈ Acij |t=0 , ζ ∈ Acki+1 |t=0 .

(4)

356

C.-Q. Cheng

'

For each integer i ∈ i j ∈Λ3 {i j , i j +1, · · · , i j+1 −1}, there exist two closed 1-forms ηi , µ¯ i defined on M, a U-step 1-form µi defined on (u,t) ∈ M × R and an open set Ui ⊂ M such that [ηi ] = ci , µi is closed on Ui × [0,t0 ], µi = 0 when t
0, µi = µ¯ i when t  t0 > 0, [µ¯ i ] = ci+1 − ci and there is a small number δi > 0 such that Cηi ,µi (t) + δi ⊂ Ui ,

when t ∈ [0,t0 ].

(5)

All orbits in C˜ηi ,µi are the local minimal orbits of the second type of L, they connect N˜ (ci ) to N˜ (ci+1 ). By the compactness of the manifold M, for a small εi∗ > 0 there exists (T˘i0 , T˘i1 ) = (T˘i0 , T˘i1 )(εi∗ ) ∈ (Z+ , Z+ ) such that ∗ hη0i ,µ1i (m0 , m1 )  h∞ ηi ,µi (m0 , m1 ) − εi T ,T

(6)

holds for all T0  Ti0 , T1  Ti1 and for all (m0 , m1 ) ∈ M × M. Obviously, given (m0 , m1 ) there are infinitely many T0  Ti0 and T1  Ti1 such that ∗ |hTη0i ,,Tµ1i (m0 , m1 ) − h∞ ηi ,µi (m0 , m1 )|
εi .

(7)

Let γi (t, m0 , m1 , T0 , T1 ) : [−T0 , T1 ] → M be the minimizer of hTη0i ,,Tµ1i (m0 , m1 ), it follows from lemma 8 that if εi∗ > 0 is sufficiently small, T0 > T˘i0 and T1 > T˘i1 are chosen sufficiently large so that (7) holds, then d γi (t, m0 , m1 , T0 , T1 ) ∈ C˜ηi ,µi (t) + δi ,

∀ 0
t
1.

(8)

T ,T By the Lipschitz property of hη0i ,µ1i (m0 , m1 ) in (m0 , m1 ) there exist Tˆi0 (εi∗ ) > T˘i0 (εi∗ ) and Tˆi1 (εi∗ ) > T˘i1 (εi∗ ) so that for each (m0 , m1 ) there are T j = T j (m0 , m1 ) with T˘i j (εi∗ )
T j
Tˆi j (εi∗ ) ( j = 0, 1) such that both (7) and (8) hold. Note that for different (m0 , m1 ) we may need different T j  T˘i j ( j = 0, 1). Before we formulate the variational principle, let us observe some fact. Let us consider these two orbits of the Lagrangian flow: one orbit of φLt , d γ : R → T M, has the property that the α -limit set α (d γ ) ⊆ A˜(c) and the ω -limit set ω (d γ ) ⊆ A˜(c ), another orbit of φLt , d γ : R → T M, has the property that the α -limit set α (d γ ) ⊆ A˜(c ) and the ω -limit set ω (d γ ) ⊆ A˜(c

). It is not necessary that dc (ω (d γ ), α (d γ )) = 0. However, under the condition II, we can use some c -semi static orbits to connect them in the sense of pseudo-metric dc (cf Theorem 9).

Proposition 17 Assume the Aubry distance from an Aubry class A0i (c) to other Aubry classes has a positive lower bound, dc (A0i (c), A0j (c))  d > 0 for all j = i, then there is an open neighborhood Nci ⊃ A0i (c), for all m0 , m1 ∈ Nci and for all x ∈ A0i (c), we have ∞ ∞ h∞ c (m0 , x) + hc (x, m1 ) = hc (m0 , m1 );

(9)

Variational methods for the problem of Arnold diffusion

357

for all m0 , m1 ∈ Nci and any x ∈ A0 (c)\A0i (c) we have d ∞ ∞ h∞ c (m0 , x) + hc (x, m1 )  hc (m0 , m1 ) + . 2

(10)

Proof. For each pair of points (m0 , m1 ) ∈ M × M, we claim that there exists some Aubry class A0j (c) such that ∞ ∞ h∞ c (m0 , m1 ) = hc (m0 , ξ ) + hc (ξ , m1 )

holds for each ξ ∈ A0j (c). Indeed, let ki → ∞ be a subsequence of integers such that lim hkci (m0 , m1 ) = h∞ c (m0 , m1 )

i→∞

let γcki : [0, ki ] → M be the minimizer for hkci (m0 , m1 ). For any large but finite number N > 0, the set {γcki |[0,N] } is compact in C1 -topology, thus we obtain forward semi static curve γcs : [0, ∞) → M. The ω -limit set of d γcs must be some Aubry class, let’s say ω (d γcs ) ⊆ A i (c). Obviously, for any ξ ∈ A i (c) the equality (9) holds. Choose a neighborhood Nci of A0i (c) such that Nci = {m ∈ M : dc (m, x)


d , ∀x ∈ A0i (c)} 6 max{1,CL }

where CL is the Lipschitz constant of the barrier function. Given m ∈ Nci , we claim that (9) and (10) hold if we let m0 = m1 = m. In fact, let ki → ∞ be a sequence such ki that limki →∞ hkci (m, m) = h∞ c (m, m) and let γm (t): [0, ki ] → M be the minimizer of hkci (m, m), the ordinary distance d(γmki (t), A0j (c))  d > 0 for all integer t ∈ [0, ki ] and j = i. Otherwise we would obtain from the property that dc (A0i (c), A0j (c))  d > 0 for all j = i and the Lipschitz property of hc (x, y) on x and y that h∞ c (m, m)  d − 2CL

3 d  d. 6 max{1,CL } 5

On the other hand, the Lipschitz property of the Barrier function Bc (x) in x induces that 2 h∞ c (m, m)
d. 5 This contradiction verifies our claim. Now we consider any two points m0 , m1 ∈ Nci . For any x ∈ A0j (c) with j = i, we let ζui (t, m0 , x): [0, ki ] → M be the curve which minimizes the quantity hkci (m0 , x) i

and limki →∞ hkci (m0 , x) = h∞ c (m0 , x), let ζs (t, x, m1 ): [0, ki ] → M be the curve which k

k

minimizes the quantity hci (x, m1 ) and limki →∞ hci (x, m1 ) = h∞ c (x, m1 ). Let  = 0, 1, ki

ki

γmi  : [0, ki ] → M be the minimizer of hc (m , m ) with limki →∞ hc (m , m ) = 

i i i i i h∞ c (m , m ), clearly ∃ x ∈ A0 (c) and integer tm ∈ [0, k ] such that γm (t ) → x

358

C.-Q. Cheng ki

i : [0, ki ] → M be the minimizer of h 01 (x , x ) and ti → ∞ as ki → ∞. Let ξ01 c 0 1 01

with limki ki

ki

01 →∞

i i hc01 (x0 , x1 ) = h∞ c (x0 , x1 ), let ξ10 : [0, k10 ] → M be the minimizer of ki

hc10 (x1 , x0 ) with limki

10 →∞

hc10 (x1 , x0 ) = h∞ c (x1 , x0 ). Given arbitrarily small δ > 0,

i and ki such that we have sufficiently large ki , ki , k0i , k1i , k01 10 ki |h∞ c (m0 , x) − hc (m0 , x)| < δ , k

i |h∞ c (x, m1 ) − hc (x, m1 )| < δ ,

ki

 |h∞ c (m , m ) − hc (m , m )| < δ ,

 = 0, 1

i k01

|h∞ c (x0 , x1 ) − hc (x0 , x1 )| < δ , ki

10 |h∞ c (x1 , x0 ) − hc (x1 , x0 )| < δ ,

Since x0 , x1 ∈ A0i (c), we have dc (x1 , x0 ) = 0. Consequently, ti

ki

ki −t1i

hc0 (m0 , x0 ) + hc01 (x0 , x1 ) + hc1

(x1 , m1 )

(11)

ki ki −t i + hc (m1 , x1 ) + hc10 (x1 , x0 ) + hc0 0 (x0 , m0 ) t1i

2
d + 6δ . 6 Since x is in another Aubry class, we have

i

i

hkc (m0 , x) + hkc (x, m1 ) i k10

t1i

(12) k0i −t0i

+ hc (m1 , x1 ) + hc (x1 , x0 ) + hc 1  d − d − 5δ . 6

(x0 , m0 )

Because δ is arbitrarily small, we obtain from (11) and (12) that 1 ∞ h∞ c (m0 , x) + hc (x, m1 ) − d 2 ∞ ∞  h∞ c (m0 , x0 ) + hc (x0 , x1 ) + hc (x1 , m1 )

 h∞ c (m0 , m1 )

it verifies (10). Since (10) holds for any point in any other Aubry class, (9) must hold for some, thus for all x ∈ A0i (c). Consider each integer i ∈

'

i j ∈Λ1 ∪Λ2 {i j , i j +1, · · · , i j+1 −1}. In view of the proposition (17), we see that for any two points m0 , m1 ∈ Ncji there exists T˘i (εi∗ ) > 0, independent of m0 and m1 , such that for all T  T˘i (εi∗ ) ∞ ∗ hTci (m0 , m1 )  h∞ ci (m0 , ζ ) + hci (ζ , m1 ) − εi ,

∀ ζ ∈ A0j (ci ),

(13)

Variational methods for the problem of Arnold diffusion

359

and there exists Tˆi (εi∗ ) > T˘i (εi∗ ) such that for each (m0 , m1 ) ∈ Ncji × Ncji we have some integers T between T˘i (εi∗ ) and Tˆi (εi∗ ) so that ∞ ∗ |hTci (m0 , m1 ) − h∞ ci (m0 , ζ ) − hci (ζ , m1 )|
εi ,

∀ ζ ∈ A0j (ci ).

(14)

Since there are finitely many Aubry class for each c, we can choose those T˘i (εi∗ ) and Tˆi (εi∗ ) which apply to each Aubry class. Let d γi be a local minimal orbit of the first or second type, connecting µci to µci+1 . The subindex for each Aubry class Acij is chosen so that A˜c1i ⊇ ω (d γi−1 ), A˜cki i ⊇ α (d γi ) and there is a ci -semi static orbit connecting A˜cij to A˜cij+1 ( j = 1, 2, · · · , ki − 1). The condition II guarantees these orbits possess the property of local minimality: there exist two open disks Vi,−j and Vi,+j+1 with V¯i,−j ⊂ (Acij |t=0 + δ )\A0 (ci ), V¯i,+j+1 ⊂ (Acij+1 |t=0 + δ )\A0 (ci ), a positive integer Ti, j and a small number εi∗ > 0 such that Ti, j ∞ min h∞ ci (ξ , m0 ) + hci (m0 , m1 ) + hci (m1 , ζ ) :

(m0 , m1 ) ∈ ∂ (Vi,−j ×Vi,+j+1 ) Ti, j ∞  min h∞ ci (ξ , m0 ) + hci (m0 , m1 ) + hci (m1 , ζ ) :

(15) (m0 , m1 ) ∈ Vi,−j ×Vi,+j+1 + 5εi∗ where ξ ∈ Acij |t=0 , ζ ∈ Acij+1 |t=0 . ' We define τi inductively for 0
i
im . Let τ0 = 0. For each i ∈ i j ∈Λ1 ∪Λ2 {i j , i j + ' 1, · · · , i j+1 − 1} with i − 1 ∈ i j ∈Λ1 ∪Λ2 {i j , i j + 1, · · · , i j+1 − 1} also, we define τi,0 , τi,1 , · · · , τi,ik and let τi = τi,0 , τi+1 = τi,ik such that 1 1 + T˘i
τi,1 − τi
T˜i−1 + Tˆi , T˜i−1

Ti, j + T˘i
τi, j − τi, j−1
Ti, j + Tˆi , T˜i0 + T˘i + Ti,ik −1




∀ j = 2, 3, · · · , ik − 1,

τi+1 − τi,ik −1
T˜i0 + Tˆi + Ti,ik −1 ,

(16) (17) (18)

1 , T˜ 0 , see (13), (14) for the definition of T˘ , Tˆ and see (2) for the definition of T˜i−1 i i i ' see (15) for the definition of Ti, j . For i ∈ i j ∈Λ3 {i j , i j + 1, · · · , i j+1 − 1} with i − 1 ∈ ' i j ∈Λ3 {i j , i j + 1, · · · , i j+1 } also, we choose those τi such that 1 1 + 1}
τi − τi−1
max{Tˆi0 , Tˆi−1 + 1}. max{T˘i0 , T˘i−1

(19)

If there is a local minimal orbit of the first or the second type d γi such that α (d γi ) ⊆ A˜(ci ) and ω (d γi ) ⊆ A˜(ci ), there is a local minimal orbit of the third type d γi+1 such that α (d γi+1 ) ⊆ A˜(ci+1 ) and ω (d γi+1 ) ⊆ A˜(ci+2 ). In this case we 0 can be taken large enough such that for any m , m ∈ M note that both Tˆi+1 and Tˆi+1 0 1 0 }
T (m , m ), T (m , m )
there exist T (m0 , m1 ), T0 (m0 , m1 ) with max{T˘i+1 , T˘i+1 0 1 0 0 1 0 } such that (7) holds if we set T = T (m , m ) there; (14) holds if we max{Tˆi+1 , Tˆi+1 0 0 0 1

360

C.-Q. Cheng

0 }. Thus, set T = T (m0 , m1 ) there; (6) and (13) hold for each T0 , T  max{T˘i+1 , T˘i+1 we choose τi,0 , τi,1 , · · · , τi,ik −1 in the way given by (16) and (17) and choose τi+1 so that 0 0 }
τi+1 − τi,i −1
Ti,i −1 + max{Tˆi , Tˆi+1 }. (20) Ti,i −1 + max{T˘i , T˘i+1 k

k

k

If both N0 (ci−1 ) and N0 (ci ) are homologically trivial and µi can be connected to µi+1 by local connecting orbits of the first or second type, we can choose suitably large Tˆi and Tˆi1 and set the range for τi,1 , τi,2 , · · · , τi,ik in the way given by (17) and (18) 1 1 , T˘i }
τi,1 − τi−1
T˜i0 + max{Tˆi−1 , Tˆi }. (21) T˜i0 + max{T˘i−1 Consider τ as the time translation τ ∗ φ (x,t) = φ (x,t + τ ) on M × R, let ψi ≡ 0 for ' i ∈ i j ∈Λ3 {i j , i j + 1, · · · , i j+1 − 1}, we define a modified Lagrangian L˜ = L − η0 −

im −1

∑ (−τi )∗ (µi + ψi ).

(22)

i=0

'

For i ∈ i j ∈Λ1 ∪Λ2 {i j , i j + 1, · · · , i j+1 − 1}, we let τ i = (τi,0 , τi,1 , · · · , τi,ik −1 ), for ' i ∈ i j ∈Λ3 {i j , i j + 1, · · · , i j+1 − 1}, we let τ i = τi . Define

τ = (τ 0 , τ 1 , · · · , τ im −1 ).

(23)

We define an index set for τ :

Λ = τ ∈ Zim −1 : (16 ∼ 21) hold .

(24)

'

± ± ± ± For i ∈ i j ∈Λ1 ∪Λ2 {i j , i j + 1, · · · , i j+1 }, we let z± i = (zi,1 , zi,2 , · · · , zi,ik ) and Vi = ± ± ,Vi,2 , · · · ,Vi,i±k ), and define (Vi,1


+ − + z = z− 0 , zi , zi , zim , i ∈

/

 {i j , i j + 1, · · · , i j+1 } ,

its domain is restricted in
− + V = V0− , V+ i , Vi ,Vim , i ∈

/

 {i j , i j + 1, · · · , i j+1 } .

i j ∈Λ1 ∪Λ2

For (m, m ) ∈ M × M and z ∈ V, we define

(m0 , m1 , z, τ ) = inf hLK,K ˜

 K +T˜ 1 +Tˆi +τi −1 m m im −1

−K im −1

+

˜ γ (t),t)dt L(d

∑ (τi − τi−1 )α (ci ) + K α (c0 ) + K α (cim )

i=1

(25)

i j ∈Λ1 ∪Λ2

(26)

Variational methods for the problem of Arnold diffusion

361

where the infimum is taken under the conditions: γ (−K) = m0 , γ (K¯ + τim −1 ) = ' m1 ; for i ∈ i j ∈Λ1 {i j , i j + 1, · · · , i j+1 − 1}, [γ |t∈[τi −T˜ 0 ,τi +T˜ 1 ] ]i = 0; for i ∈ i i ' − + ˜0 ˜1 ; γ (τi, j ) = z− i ∈Λ ∪Λ {i j , i j +1, · · · , i j+1 −1}, γ (τi − Ti ) = zi,i , γ (τi + Ti ) = z i, j j

1

2

i+1,1

k

and γ (τi, j + Ti, j ) = z+ i, j+1 for each j = 1, 2, · · · , ik − 1.



Let hLK,K (m0 , m1 ) be the minimizer of hLK,K (m0 , m1 , z, τ ) over V in z and over Λ ˜ ˜ in τ respectively:



hLK,K (m0 , m1 ) = min hLK,K (m0 , m1 , z, τ ), ˜ ˜ τ ∈Λ ,z∈V

let K j , K j → ∞ be the subsequence such that K j ,K j

hL˜ lim

K j ,K j →∞

(m0 , m1 ) = lim inf hLK,K (m0 , m1 ), ˜

K→∞K →∞

and denote the corresponding minimal curve by γ (t; K j , K j , m0 , m1 ), we claim that d γ (t; K j , K j , m0 , m1 ) is a solution of the Euler–Lagrange equation determined by L large. Indeed, if K j and K j are sufficiently ' 1. For each i ∈ i j ∈Λ3 {i j , i j + 1, · · · , i j+1 − 1}, we have (−τi )∗ γ (t; K j , K j , m0 , m1 ) ∈ Cηi ,µi (t) + δi ⊂ Ui ,

(27)

when τi
t
τi + 1. To see it, let us choose mi = γ (τi−1 + 1), m i = γ (τi+1 ). Since

(m0 , m1 , Z, τ ) over Λ , thus the curve γ (t; K j , K j , m0 , m1 ) is the minimizer of hLK,K ˜ τ

AL˜ ((−τi )∗ γ |τi+1 ) + (τi − τi−1 + 1)α (ci ) + (τi+1 − τi α (ci+1 ) i−1 +1 =

inf

ξ (−T0 )=mi ξ (T1 )=m i T˘i0
T0
Tˆi0 T˘i1
T1
Tˆi1

 T1

−T0

(L − ηi − µi )(d ξ (t),t)dt + T0 α (ci ) + T1 α (ci+1 ).

Thus we obtain (27) from this formula, (5), (8) and (19). Consequently, γ (t; K j , K j )|τi
t
τi +1 falls into the region where (−τi )∗ µi is closed. So, d γ (t; K j , K j , m0 , m1 ) is the solution of the Euler–Lagrange equation determined by L when τi
t
τi + 1; ' 2. For i ∈ i j ∈Λ1 {i j , i j + 1, · · · , i j+1 − 1}, we claim that (−τi )∗ γ (t)|0
t
t0 ∈ int(Oi ).

(28)

˜ C˜ηi ,µi ,ψi (M) then γ must It is the consequence of (1). In fact, if d γ ∈ π1 C˜ηi ,µi ,ψi (M)\ pass through Oi during the time interval [0,t0 ]. Note, the function Lηi ,µi ,ψi is no ˜ C˜ηi ,µi ,ψi (M), k∗ d γi is not a minimizer longer time-periodic, if d γi ∈ π1 C˜ηi ,µi ,ψi (M)\ ˜ t=0 for each k ∈ Z\{0}. / π1 C˜ηi ,µi ,ψi (M)| of this kind, d γi (k) ∈

362

C.-Q. Cheng

˜ C˜ηi ,µi ,ψi (M), we assume that α (d γ ) ⊂ Given a smooth curve d γ ∈ C˜ηi ,µi ,ψi (M)\ ik i 1 ˜ ˜ Aci and ω (d γ ) ⊂ Aci+1 . For any mi ∈ Mcik |t=0 , mi+1 ∈ Mc1i+1 |t=0 and , if Z+  T0k → ∞ and Z+  T1k → ∞ (as k → ∞) are two sequences such that γ (−T0k ) → mi and γ (T1k ) → mi+1 , then lim

 Tk 1

k→∞ −T k 0

Lηi ,µi ,ψi (d γ (t),t)dt + T0k α (ci ) + T1k α (ci+1 ) = h∞ ηi ,µi ,ψi ,e1 (mi , mi+1 ).

Let ζ : R → M be an absolutely continuous curve such that [ζ ]i = 0, ζ (t ) ∈ / int(Oi ) for some t ∈ [0,t0 ], ζ (−T0k ) → mi , ζ (T1k ) → mi+1 as k → ∞. Since C˜ηi ,µi ,ψi |t=constant is closed, there exists a positive number d > 0 such that lim inf T0k →∞ T1k →∞

 Tk 1 −T0k

Lηi ,µi ,ψi (d ζ (t),t)dt + T0k α (ci ) + T1k α (ci+1 )  h∞ ηi ,µi ,ψi ,e1 (mi , mi+1 ) + d.

Recall the construction of the modified Lagrangian L˜ (see (22)) and γ is the mini mizer of hLK,K (m0 , m1 , z, τ ) over V in z and over Λ in τ respectively. Given any small ˜ number ε > 0, by choosing sufficiently large Tˆi − T˘i , we can see that there are suffi1 + K − + K + = τ − T˜ 0 , ciently large Ki− , Ki+ ∈ Z with the properties that τi−1 + T˜i−1 i i i i T˘i
Ki− + Ki+
Tˆi and γ (τi − T˜i0 − Ki+ ) − mi  < ε ,

− γi (τi + T˜i1 + Ki+1 ) − mi+1 
ε .

If there was t ∈ [0,t0 ] such that / int(Oi ), (−τi )∗ γ (t ) ∈

from the Lipschitz continuity of h∞ ηi ,µi ,ψi ,ei (m, m ) in (m, m ) we would obtain

 τi +T˜ 1 +K − i i+1 τi −T˜i0 −Ki+

+ − Lηi ,µi ,ψi (d γ (t),t)dt + (T˜i0 + Ki−1 )α (ci ) + (T˜i1 + Ki+1 )α (ci+1 )

3  h∞ ηi ,µi ,ψi ,e1 (mi , mi+1 ) + d. 4 On the other hand, there is suitably large K¯ i− , K¯ i+ ∈ Z with the properties that T˘i
K¯ i− + K¯ i+
Tˆi and γi (τi − T˜i0 − K¯ i+ ) − mi 
ε , Because d γi (t ) ∈ C˜ηi ,µi ,ψi |t=t , we have

− γi (τi + T˜i1 + K¯ i+1 ) − mi+1 
ε .

Variational methods for the problem of Arnold diffusion

 τi +T˜ 1 +K¯ − i i+1 τi −T˜i0 −K¯ i+

363

+ − Lηi ,µi ,ψi (d γi (t),t)dt + (T˜i0 + K¯ i−1 )α (ci ) + (T˜i1 + K¯ i+1 )α (ci+1 )

1
h∞ ηi ,µi ,ψi ,ei (mi , mi+1 ) + d 4 if ε is sufficiently small. It implies that γ is not a minimizer. This contradiction verifies our claim. The formula (28) implies that d'γ (t; K j , K j ) is the solution of the Euler–Lagrange equation determined by L for i ∈ i j ∈Λ1 {i j , i j + 1, · · · , i j+1 − 1}. ' 3. For i ∈ i j ∈Λ2 {i j , i j + 1, · · · , i j+1 − 1}, we claim that (28) also holds. In this case, we have (3) instead of (1). The argument is the same if we replace the quantity ∞ h∞ ηi ,µi ,ψi ,ei (m0 , m1 ) by hηi ,µi ,ψi (m0 , m1 ). 4. We claim that the curve γ does not touch the boundary of Vi,i−k at the time + t = τi − T˜i0 and does not touch the boundary of Vi+1,1 at the time t = τi + T˜i1 for each ' + integer i ∈ i j ∈Λ1 {i j , i j +1, · · · , i j+1 −1}. If (γ (τi − T˜i0 ), γ (τi + T˜i1 )) = (z− i,ik , zi+1,1 ) ∈ − − + + ∂ (Vi,ik × Vi+1,1 ), let zi,ik = γ (τi−1,(i−1)k −1 + Ti−1,(i−1)k −1 ) and zi+1,1 = γ (τi,1 ), from − + + k (2) we can see that there exist (¯z− i,ik , z¯i+1,1 ) ∈ Vi,ik ×Vi+1,1 such that for ξ ∈ Mci |t=0 , ζ ∈ Mc1i+1 |t=0 : i

T˜ 0 ,T˜ 1

− − + − i+1 + i i hTcii (z+ i,ik , zi,ik ) + hηi ,µi ,ψi ,ei (zi,ik , zi+1,1 ) + hci+1 (zi+1,1 , zi+1,1 ) T

T˜ 0 ,T˜ 1

− − + ∞ + i i  h∞ ci (ξ , zi,ik ) + hηi ,µi ,ψi ,ei (zi,ik , zi+1,1 ) + hci+1 (zi+1,1 , ζ ) + ∞ − ∗ + h∞ ci (zi,ik , ξ ) + hci+1 (ζ , zi+1,1 ) − 2εi T˜ 0 ,T˜ 1

− + ∞ i i  h∞ z− z+ ci (ξ , z¯i,ik ) + hηi ,µi ,ψi ,ei (¯ i,ik , z¯i+1,1 ) + hci+1 (¯ i+1,1 , ζ ) + ∞ − ∗ + h∞ ci (zi,ik , ξ ) + hci+1 (ζ , zi+1,1 ) + 3εi T˜ 0 ,T˜ 1

− + ∗ i+1 + i i  hTcii (z+ z− zi+1,1 , z− i,ik , z¯i,ik ) + hηi ,µi ,ψi ,ei (¯ i,ik , z¯i+1,1 ) + hci+1 (¯ i+1,1 ) + εi T

satisfy the condition T˘
T , T
Tˆ ( j = i − 1, i). In above where Ti , Ti+1 , Ti , Ti+1 j j j j arguments, (13) and (14) are used to obtain the first and the third inequality, (2) is used to obtain the second inequality. But this'contradicts to the fact that γ is a minimal curve of L˜ on V and Λ . The case for i ∈ i j ∈Λ2 {i j , i j + 1, · · · , i j+1 − 1} can be treated in the same way. Therefore, the minimizer γ is differentiable at the time t = τi − T˜i0 and t = τi + T˜i1 for each i. 5. We claim that the curve γ does not touch the boundary of Vi,−j at the time t = τi, j and does not touch the boundary of Vi,+j+1 at the time t = τi, j + Ti, j for each integer ' i ∈ i j ∈Λ1 ∪Λ2 {i j , i j + 1, · · · , i j+1 − 1} and j ∈ {1, 2, · · · , ik − 1}. The demonstration is similar to the case 4, based on (13), (14) and (15). Let K j , K j → ∞ and let γ∞ : R → M be an accumulation point of the curves {γ (t, K j , K j )}. Obviously, α (d γ∞ ) ⊂ A˜(c) and ω (d γ∞ ) ⊂ A˜(c ). This completes the proof.

364

C.-Q. Cheng

The Theorem 10 provides us a possible way to prove Arnold diffusion is a generic phenomenon for positive definite systems. However, the verification of the condition I is difficult, when the conditions II and III do not hold. It contains two two key points: for each c in the chain the Aubry set does not generate the first homology group H1 (M, A0 (c), Z) = 0 and the barrier function is not constant when it is restricted outside of the Aubry set. In general case, we are unable to show the genericity of these conditions, however, it can be down in some interesting cases.

6 Application to a priori unstable systems Here we study a typical example of a priori unstable system: H(I, φ , x, y,t) = h0 (I) + h1 (x, y) + ε P(I, φ , x, y,t), where (I, φ ) ∈ R × T, (x, y,t) ∈ Tn × Rn × T. Let ˙ x,t) = 0 (φ˙ ) + 1 (x, ˙ x) + ε L1 (φ˙ , φ , x, ˙ x,t) L(φ˙ , φ , x, be the Legendre transformation of the Hamiltonian. We call it priori unstable when (x, ˙ x) = (0, 0) is a hyperbolic fixed point which corresponds to the minimum of the action of 1 . Under the a priori unstable condition, the time-1-map Φ of the Hamiltonian flow has an invariant cylinder which is a small deformation of the standard cylinder Σ = {(I, φ ) ∈ R × T}. The restriction of Φ on Σ is area-preserving and twist. In virtue of many works before, we have very well understanding on the dynamics on the cylinder. Under Legendre transformation, this cylinder and the dynamics have their correspondence in the space of tangent bundle. Without danger of confusion, we use the same name for the object and its correspondence under the Legendre transformation. Let Γ : [0, 1] → H 1 (Tn+1 , R) be a path such that Γ (s) = (c1 (s), 0, · · · , 0) where the first component represents for φ˙ . As the hyperbolic property is assumed, for each s ∈ [0, 1], the Aubry set is in on the cylinder and the support of the minimal measure is either invariant curve, or the Aubry–Mather set, or the minimal periodic orbit. Therefore, the condition H1 (M, A0 (c), Z) = 0 is satisfied. To verify the condition: ˜ (I) There is small δτ > 0 such that π1 N0 (Γ (τ ), M)\(A 0 (Γ (τ ), M) + δτ ) is nonempty and totally disconnected We need to study some regularity of the barrier function on c. We find that it is Hölder continuous with exponent 12 at the points where there is an invariant curve. Consequently, we are able to show this condition is also generic (cf. [CY2]). Therefore, Arnold diffusion is a generic phenomenon in a priori unstable systems.

Variational methods for the problem of Arnold diffusion

365

References [Ar1]

[Be1] [Be2] [Be3] [CY1] [CY2] [CDI] [CP] [Fa1] [Fa2] [Ma1] [Ma2] [Me2]

Arnol’d V. I., Instability of dynamical systems with several degrees of freedom, (Russian, English) Sov. Math., Dokl., 5(1964), 581–585; translation from Dokl. Akad. Nauk SSSR, 156(1964), 9–12. Bernard P., Homoclinic orbits to invariant sets of quasi-integrable exact maps, Ergod. Theory Dynam. Syst. 20(2000), 1583–1601. Bernard P., Connecting orbits of time dependent Lagrangian systems, Ann. Inst. Fourier, 52(2002), 1533–1568. Bernard P., The dynamics of pseudographs in convex Hamiltonian systems, preprint (2004) Cheng C. -Q. & Yan J., Existence of diffusion orbits in a priori unstable Hamiltonian systems, J. Differential Geometry, 67(2004), 457–517. Cheng C. -Q. & Yan J., Arnold diffusion in Hamiltonian systems: a priori unstable case, preprint (2004). Contreras G., Delgado J. & Iturriaga R., Lagrangian flows: the dynamics of globally minimizing orbits II, Bol. Soc. Bras. Mat. 28(1997), 155–196. Contreras G. & Paternain G. P., Connecting orbits between static classes for generic Lagrangian systems, Topology, 41(2002), 645–666. Fathi A., Théorème KAM faible et théorie de Mather sue les systèmes, C. R. Acad. Sci. Paris Sér. I Math. 324(1997), 1043–1046. Fathi A., Weak KAM Theorem in Lagrangian Dynamics (to appear) Cambridge University Press. Mather J., Action minimizing invariant measures for positive definite Lagrangian systems, Math. Z., 207(2)(1991), 169–207. Mather J., Variational construction of connecting orbits, Ann. Inst. Fourier (Grenoble), 43(5)(1993), 1349–1386. Mañé R., Generic properties and problems of minimizing measures of Lagrangian systems, Nonlinearity, 9(1996), 169–207.

The calculus of variations and the forced pendulum Paul H. Rabinowitz1

Abstract Consider the equation of forced pendulum type: u

+Vu (t, u) = 0 (∗) where = d/dt and V is smooth and 1-periodic in its arguments. We will show how to use elementary minimization arguments to find a variety of solutions of (∗). We begin with periodic solutions of (∗) and then find heteroclinic solutions making one transition between a pair of periodics. Then we construct heteroclinics and homoclinics making multiple (even infinitely many) transitions between periodics. If time permits, we may also discuss the construction of related mountain pass orbits of (∗).

1 Introduction The goal of these lectures is to show how elementary variational techniques, in particular minimization arguments, can be used to extract a considerable amount of information about dynamical behavior. We do this for the setting of a forced pendulum model problem. This is a favorite proving ground for many techniques. Among works that are related to ours, we mention in particular [Mor], [A], [Ma82], [Ma93], [B88], [B89], and [Mos86]. The approach taken here uses essentially nothing from the theory of dynamical systems other than the uniqueness of solutions of the initial value problem. Therefore, these techniques can also be used for certain classes of problems for partial differential equations. In part our arguments are simplifications of ones used in [RS]. A disadvantage of our approach is that it does not capture finer dynamical structure that can be obtained using stable and unstable manifolds or notions like hyperbolicity. 1

University of Wisconsin-Madison, Mathematics Department, 480 Lincon Dr, Madison WI 53706-1388 e-mail: [email protected]

W. Craig (ed.), Hamiltonian Dynamical Systems and Applications, 367–390. c 2008 Springer Science + Business Media B.V.


367

368

P.H. Rabinowitz

Fig. 1 Schematic of the physical pendulum.

Fig. 2 Schematic of an orbit asymptotic from v to w.

Fig. 3 Schematic of an orbit asymptotic from w to v.

The simple pendulum is modeled by u

+sin u = 0, u representing the angle made with the vertical direction. More generally we will consider a forced model (DE) − u

+Vu (t, u) = 0, where V satisfies (V1 ) V ∈ C2 (R2 , R) and is 1-periodic in t and in u. Equivalently V ∈ C2 (T2 , R), where T2 is the 2-torus. A caveat is in order here: V is the negative of the usual potential energy. The simplest solutions of (DE) are periodic ones, e.g. if Vu (t, z) = 0 for all t ∈ R and z ∈ Z, each such z is an equilibrium, and therefore periodic solution of (DE). By (V1 ), if v is a solution of (DE), so is v + k for all k ∈ Z. Therefore we can seek solutions of (DE) that are asymptotic to a pair of periodics v and w. We say such a solution is heteroclinic from v to w (Fig. 4). Such solutions undergo one ‘transition’. Likewise we can try to find 2, k or infinite transition solutions. Thus a 2-transition solution is homoclinic to v or w (see Figs. 2–4). It turns out there are infinitely many solutions of each type, distinguished by the amount of time they spend near v or w.

The calculus of variations and the forced pendulum

369 w u u v

Fig. 4 Graphs of 1-transition orbits between v and w. w u v

Fig. 5 Graphs of 2-transition orbits between v and w. v+2 u

v+1 v

Fig. 6 A monotonic orbit asymptotic to v in the past and to v + 2 in the future. v+3

u

v+2 v+1 v

Fig. 7 An orbit which makes several transitions.

There is another kind of 2-transition solution which is monotone: u(t + 1) > u(t) (Fig. 1). In the simplest case, such a solution is heteroclinic from v to v + 2. Likewise, there are k and infinite-translation such solutions, and we can concatenate these two types of solutions (Fig. 7). Within each type of solution as well as for the mixed type, one can seek a socalled symbolic dynamics of solutions that will be described later. We will show how elementary minimization arguments can be used to find some of these solutions. Unfortunately we will not have enough time to treat the monotone and mixed cases. We begin with the simplest case of periodic solutions and then treat progressively more complex cases.

2 Periodic solutions Periodic solutions are the easiest to find. We assume V satisfies (V1 ). Set E = W 1,2 (T1 ), the class of 1-periodic functions having square integrable derivatives, i.e. 2 ||u||2E = ||u||W 1,2 =

 1  (u )2 + u2 dt. 0

370

P.H. Rabinowitz

Note that u ∈ E implies u ∈ C(T1 ), in fact u ∈ C1/2 (T1 ), i.e. u is Hölder continuous of order 1/2. Let 1 L(u) = |u |2 +V (t, u), 2 be the Lagrangian associated with (DE) with the corresponding functional  1

I(u) =

L(u) dt. 0

Then I ∈ C(E, R) (even C2 ) and for u, φ ∈ E, the Frechet derivative, I (u)φ is given by 1 I (u)φ = lim (I(u + hφ ) − I(u)) h→0 h  1  u φ +Vu (t, u)φ dt. = 0

I (u) = 0,

If we say u is a critical point of I and c = I(u) is called a critical value of I. Note also, if  1  u φ +Vu (t, u)φ dt = 0 (1) 0

for all φ ∈ E, u is called a weak solution of (DE). Then we have a “regularity” theorem: Theorem 2.1. u is a classical solution of (DE) if and only if u ∈ E and u is a weak solution of (DE). Theorem 2.1 reduces the existence of periodic solutions of (DE) to finding critical points of I in E. In the study of partial differential equations, such regularity theorems are often rather delicate. For the above special case, the proof is quite direct. Since the regularity question will also come up in more complicated settings later, we treat it here for the simplest case. Proof of Theorem 2.1. If u is a classical solution of (DE), multiplying (DE) by φ ∈ E and integrating over [0, 1] yields (1). Conversely suppose u is a weak solution of (1). Taking φ = 1 shows  1 0

V (t, u)dt ≡ [V (t, u)] = 0,

i.e. the constant term in the Fourier expansion of V (t, u) vanishes. It is a calculus exercise to show there is a unique q ∈ C2 (T1 , R) solving −q¨ +Vu (t, u) = 0 , [q] = 0.

(2)

Multiplying (2) by φ ∈ E and integrating over [0, 1] shows  1  q φ +Vq (t, u)φ dt = 0. 0

(3)

The calculus of variations and the forced pendulum

371

Subtracting (3) from (1) gives  1 0

 u − q φ dt = 0

(4)

for all φ ∈ E. Choosing φ = u − q, (4) implies u − q = 0 and therefore u = q +  const ∈ C2 (T1 , R).

How do we find critical points of I? The simplest possibilities are minima. Thus set c = inf I(u).

(5)

u∈E

Note that I is bounded from below by V0 = minR2 V . Let (un ) be a minimizing sequence for (5), i.e. I(un ) → c as n → ∞. Therefore there is an M > 0 such that   1 1 2 (u ) +V (t, un ) dt ≤ M. I(un ) = 2 n 0 Hence

||u n ||2L2 ≤ 2(M −V0 ).

(6)

Observe that un + jn is also a minimizing sequence for (5) for any choice of jn ∈ Z. Therefore un may not be bounded. But we can choose jn so that [un + jn ] ∈ [0, 1]. Thus without loss of generality, [un ] ∈ [0, 1). Since un (t) − un (x) =

 t x

 1  t

one has un (t) = [un ] + and therefore |un (t)| ≤ 1 +

 1 0

0

x

u n (s)ds, 

u n (s)ds

dx,

||u n ||L2 dx = 1 + ||u n ||L2 .

(7)

Now (6) and (7) show un is bounded in the Hilbert space E. Therefore there is a v ∈ E such that along a subsequence, un  v (i.e. weakly in E). The functional I is weakly lower semicontinuous. Hence c ≤ I(v) ≤ lim I(un ) = inf I = c. n→∞

E

(8)

Thus (8) shows I(v) = c and v minimizes I over E. Moreover v is a critical point of I on E. Indeed take φ ∈ E. Then ψ (h) ≡ I(v + hφ ) ∈ C1 (R, R) and has a minimum at h = 0. Hence ψ (0) = 0 = I (v)φ (9) for all φ ∈ E. Thus v is a weak and therefore by Theorem 2.1, a classical solution of (DE).

372

P.H. Rabinowitz

As was noted above, the minimizing sequence {un } is bounded in E and therefore in C1/2 (T1 ). Hence the subsequence {un } can be assumed to converge to v in L∞ (T1 ). Although it is not important here, for future reference, we have a stronger form of convergence: Proposition 2.1. un → v in E (i.e. in W 1,2 (T1 )). Proof. If not there is a δ > 0 such that ||u n − v ||L2 ≥ δ . Set φn = un − v. Then I(un ) = I(v + φn )   1 1 2 1 = |v | + v φn + |φn |2 +V (t, v + φn ) −V (t, v) +V (t, v) dt 2 2 0  1  1 v φn +V (t, v + φn ) −V (t, v) dt. (10) ≥ I(v) + δ 2 + 2 0 As n → ∞, I(un ) → I(v) while the term on the right in (10) approaches zero. Thus  0 ≥ 1/2δ 2 , a contradiction.

Set M0 = {u ∈ E : I(u) = c}. We have shown M0 = 0. / Example 1: If V ≡ 0, then M0 = R. Example 2: If V = a(t)(cos(2π u − 1)), then M0 = Z. Theorem 2.2. M0 is an ordered set, i.e. v, w ∈ M0 implies v ≡ w, v < w, or v > w. Proof. If not, there are points ξ , η ∈ [0, 1] such that v(ξ ) = w(ξ ) and, e.g. v(η ) < w(η ). Set φ = max(v, w) and ψ = min(v, w). Then φ , ψ ∈ E and 2c ≤ I(φ ) + I(ψ ) = I(v) + I(w) = 2c.

(11)

Hence by (11), I(φ ) = c = I(ψ ) and φ , ψ ∈ M0 . Consequently by Theorem 2.1, φ and ψ are classical 1-periodic solutions of (DE). Set χ = φ − ψ so χ ≥ 0, χ (ξ ) = 0 and therefore χ (ξ ) = 0, and χ (η ) > 0. (DE) implies

χ

+Vu (t, φ ) −Vu (t, ψ ) = 0 = χ

+ f (t)χ , &

where f (t) =

Vu (t,φ (t))−Vu (t,ψ (t) ψ (t)−φ (t)

Vuu (t, φ (t))

(12)

if φ (t) > ψ (t) if φ (t) = ψ (t)

and f ∈ C(T1 , R). Thus χ is a C2 solution of the linear equation (12) with χ (ξ ) = 0 = χ (ξ ). Therefore the uniqueness of solutions to the initial value problem for  (12) implies χ ≡ 0, contrary to χ (η ) > 0. Hence M0 is ordered.

Next let k ∈ Z. Note that V is k-periodic in t so we can seek k-periodic solutions of (DE). Let u ∈ W 1,2 (kT1 ) ≡ Ek . Set  k

Ik (u) =

L(u) dt, 0

The calculus of variations and the forced pendulum

373

and

αk = inf Ik (u). u∈Ek

By our above arguments, / Mk ≡ {u ∈ Ek : Ik (u) = αk } = 0, any u ∈ Mk is a classical k-periodic solution of (DE), and Mk is an ordered set. Surprisingly we gain nothing new by varying k as the next result shows: Proposition 2.2. M0 = Mk and αk = kc. Proof. Let v ∈ Mk . Then v(· + 1) ∈ Mk . If v(t) = v(t + 1) for all v ∈ Mk , then Mk = M0 and αk = kc. Otherwise for some v ∈ Mk , (a) v(t + 1) < v(t), or (b) v(t + 1) > v(t). If (a) occurs, v(t) = v(t + k) < · · · < v(t + 1) < v(t), a contradiction, and similarly for (b).

 Proposition 2.2 can be used to show that the members of M0 possess another minimality property. Proposition 2.3. Let v ∈ M0 and a, b ∈ R with a < b. Set A = {w ∈ W 1,2 [a, b] : w(a) = v(a), w(b) = v(b)} and for w ∈ A, let I (w) =

b a

L(w) dt. Then

I (v) = inf I (w) ≡ cA .

(13)

w∈A

Proof. I is weakly lower semi-continuous so as earlier, there is a u ∈ A such that I (u) = cA . Choose α < a, and β > b with α , β ∈ Z. Extend u to [α , β ] via u = v in ' [α , a] [b, β ] and further extend u to R as a β − α periodic function. Hence u ∈ Eβ −α so by Proposition 2.2, (14) Iβ −α (v) ≤ Iβ −α (u). But  a

Iβ −α (v) = ≥

α  a α

L(v) dt + I (v) + L(u) dt + I (u) +

= Iβ −α (u). Thus by (14)–(15), I (v) = I (u) = cA .

 β

L(v) dt b

 β

L(u) dt b

(15) 

374

P.H. Rabinowitz

Remark: The minimization problem (13) is a special case of inf I (w)

w∈B

(16)

where B = {w ∈ W 1,2 [a, b] : w(a) = r, w(b) = s}. By the argument of (5)–(9), problem (16) has a minimum which is a classical solution of (DE). In several future arguments we will use this observation to establish that the minimizers of certain variational problems are in fact classical solutions of (DE). Returning to M0 , since it is ordered, either {(t, u(t)) |t ∈ R, u ∈ M0 } = R2 , i.e. M0 foliates R2 , or there are points (x, z) ∈ R2 such that z = u(x) for any u ∈ M0 , i.e. M0 merely laminates R2 . In this latter case there is a smallest w ∈ M0 and largest v ∈ M0 such that v(x) < z < w(x). Hence by Theorem 2.2, v(t) < w(t) for all t ∈ R. We then call v and w a gap pair. It is known that this latter case is generic; indeed given any v ∈ M0 , there is a W ∈ C2 (T1 , R) such that the M0 associated with V + ε W is {v + k | k ∈ Z} for all small ε > 0 [RS].

3 Heteroclinic solutions Suppose v, w ∈ M0 are a gap pair. We seek solutions of (DE) that are heteroclinic from vto w (or from w to v). A natural approach is to try to find them as minimizers of R L(u) dt over a class of functions having the desired asymptotic behavior.   However if 01 L(v) dt = c = 01 L(w) dt = 0, then for each admissible function u,  R L(u) dt will be infinite. Thus this approach must be modified. The above functional must be “renormalized” so that it is finite on the above class of functions. This can be done merely assuming (V1 ), but it is technically simpler to assume V is also time reversible. Hence suppose (V2 ) V (−t, z) = V (t, z) for all t, z ∈ R A key consequence of (V2 ) is: Proposition 3.1. If V satisfies (V1 ) and (V2 ), cˆ ≡

inf

u∈W 1,2 [0,1]

I(u) = c

(1)

and if u ∈ M0 , then u(t) = u(−t). ˆ = {u ∈ W 1,2 [0, 1] : I(u) = c}. Proof. Set M ˆ The existence argument of the previous ˆ = 0. section implies M / Clearly cˆ ≤ c. To get equality, let u ∈ W 1,2 [0, 1]. Then  1/2

 1

L(u) dt +

I(u) = 0

1/2

L(u) dt ≡ α + β .

The calculus of variations and the forced pendulum

375

Say α ≤ β . Define φ (t) = u(t) for 0 ≤ t ≤ 1/2, and φ (t) = u(1 −t) for 1/2 ≤ t ≤ 1. Then φ (0) = φ (1) so φ extends naturally to an element of E and by (V2 ), I(φ ) = 2α ≤ I(u). Therefore c = inf I ≤ inf = cˆ E

W 1,2 [0,1]

ˆ But if u ∈ M, ˆ then I(φ ) = c so φ ∈ M0 . Since φ ≡ u so c = cˆ and M0 ⊂ M. on [0, 1/2], uniqueness of solutions of the initial value problem for (DE) implies u ≡ φ on [0, 1], i.e. u ∈ M0 . Moreover u(t) = u(1 − t) = u(−t) via the 1-periodicity of u.

 With the aid of Proposition 3.1, a renormalized functional can be introduced. For 1,2 (R, R), define p ∈ Z and u ∈ Wloc a p (u) ≡

 p+1 p

L(u) dt − c.

By Proposition 3.1, a p (u) ≥ 0 for all such p and u. Now we define the renormalized functional: J(u) = ∑ a p (u). p∈Z

Thus J(u) ≥ 0. With v, w a gap pair, we define,

Γ−∞ ≡ Γ−∞ (v, w) 1,2 ≡ {u ∈ Wloc (R, R) : ||u − v||L2 [i,i+1] → 0, i → −∞}

Γ∞ ≡ Γ∞ (v, w) 1,2 ≡ {u ∈ Wloc (R, R) : ||u − w||L2 [i,i+1] → 0, i → ∞}

and take as the associated class of admissible functions 1,2 Γ1 ≡ Γ1 (v, w) ≡ {u ∈ Wloc (R, R) : v ≤ u ≤ w} ∩ Γ−∞ ∩ Γ∞

Clearly Γ1 = 0/ and there are u’s in Γ1 such that J(u) < ∞. Define c1 ≡ c1 (v, w) ≡ inf J(u). u∈Γ1

Then we have Theorem 3.1. If V satisfies (V1 ) - (V2 ), and v, w are a gap pair, then / 1. M1 ≡ M1 (v, w) ≡ {u ∈ Γ1 : J(u) = c1 } = 0. 2. Any U ∈ M1 is also a classical solution of (DE). 3. u < U < U(· + 1) < w. 4. M1 is an ordered set. 5. Any U ∈ M1 is minimal in the sense of Proposition 2.3.

(2)

376

P.H. Rabinowitz

Proof. Let {uk } be a minimizing sequence for (2). Since we are dealing with an unbounded domain, some extra care must be taken here to ensure that {uk } has a nontrivial limit. E.g. if M1 = 0/ and U ∈ M1 , uk = U(· − k) ∈ Γ1 and uk converges 2 to v ∈ / Γ1 . To avoid such complications, {uk } will be normalized as follows. in Cloc → v in If u ∈ Γ1 so is u(· − l) for any l ∈ Z and J(u(· − l)) = J(u). As l → −∞, u|l+1 l l+1 2 2 L and as l → ∞, u|l → w in L . Therefore there is a unique l = l(u) ∈ Z such that $  i+1



(u(t − l) − v(t)) dt < 12 01 (w − v) dt, i < 0, i ∈ Z 1 1 0 (u(t − l) − v(t)) dt ≥ 2 0 (w − v) dt.

i1

(3)

Thus without loss of generality, {uk } can be chosen so that l(uk ) = 0. Since {uk } is a minimizing sequence, there is an M > 0 such that for all k ∈ N, J(uk ) ≤ M.

(4)

Hence for all p ∈ N, p

 p+1

−p

−p

∑ ai (uk ) = and (5) implies

 p+1 −p

L(uk ) dt − (2p + 1)c ≤ M

(5)

|u k |2 dt ≤ M1 ,

(6)

1,2 (R, R). where M1 depends on p but not k. Since v ≤ uk ≤ w, {uk } is bounded in Wloc 1,2 Consequently there is a U ∈ Wloc such that along a subsequence uk → U weakly in 1,2 ∞ . (In fact in the spirit of Proposition 2.1, u → U in W 1,2 along a Wloc and in Lloc k loc  p+1 subsequence, but we do not need this additional information). Since −p L(u) dt is weakly lower semi-continuous, p

∑

ai (U) ≤ M

i=−p

for all p ∈ N and hence J(U) ≤ M. Moreover by (3), $  i+1

 (U − v) dt ≤ 12 01 (w − v) dt, i < 0, i i 1 1 1 0 (U − v) dt ≥ 2 0 (w − v) dt.

∈Z

(7)

∞ convergence of {u } implies v ≤ U ≤ w. Thus we We claim U ∈ Γ1 . The Lloc k need only show U satisfies the asymptotic requirements of Γ1 . To do so, note first  that since J(U) < ∞, a p (U) → 0 as |p| → ∞, i.e. pp+1 L(U) dt → c as |p| → ∞. Set Up (t) = U(t + p) for t ∈ [0, 1]. Then U p ∈ W 1,2 [0, 1] and I(U p ) → c as |p| → ∞. Hence as |p| → ∞, {U p } is a minimizing sequence for (1). Consequently along a subsequence {Up } converges weakly in W 1,2 and strongly in L∞ to u± ∈ M0 . But v ≤ Up ≤ w implies either u± = v or u± = w. By (7), as p → −∞,

The calculus of variations and the forced pendulum

 1 1

2

0

(w − v) dt ≥

 p+1 p

(U − v) dt =

377

 1 0

(Up − v) dt →

 1 0

(u− − v) dt.

Therefore u− = v and since v is the only possible limit of a subsequence of {U p } as p → −∞, the full sequence U p → v as p → −∞. It remains to prove that Up → w as p → ∞. For this, we no longer have (7) to help as for p → −∞, so more work is required. Following the argument of Proposition 2.1, we can assume U p → u+ in W 1,2 [0, 1] along our subsequence. In fact, Up → u+ along the full sequence as p → ∞ for otherwise there are a pair of subsequences such that U p → v in W 1,2 along the first, and U p → w in W 1,2 along the second as p → ∞. But Up cannot only be close (in W 1,2 [0, 1] and therefore in L∞ [0, 1]) to both v and w. Therefore there is an ε > 0 and a third subsequence such that along it, ||Up − φ ||W 1,2 [0,1] ≥ ε as p → ∞ for φ = v and φ = w. Now we have Lemma 3.1. For any ε > 0, there is a γ (ε ) > 0 such that ||Up − φ ||W 1,2 [0,1] ≥ ε implies I(Up ) ≥ c + γ (ε ) Proof. Otherwise, there is a sequence of p’s going to infinity such that I(Up ) → c while ||Up − φ ||W 1,2 [0,1] ≥ ε . As above along a subsequence, Up → v or w in  W 1,2 [0, 1], a contradiction.

Completion of the Proof of Theorem 3.1. Let S = {p ∈ N : ||U p − u+ ||W 1,2 [0, 1] ≥ ε }. Then by Lemma 3.1, J(U) ≥

∑ a p (U) ≥ ∑ γ (ε ) = ∞,

p∈S

p∈S

contrary to J(U) ≤ M. Thus Up → u+ in W 1,2 [0, 1] as p → ∞. Now finally to show that u+ = w, suppose u+ = v. By the reasoning just used and (7), there is an i ∈ Z, i ≤ 0, and ε > 0 such that ||Ui − φ ||W 1,2 [0,1] ≥ ε with φ = v and φ = w. Hence by Lemma 3.1 ai (U) ≥ γ (ε ). Therefore for large k, 1 ai (uk ) ≥ γ (ε ). 2

(8)

Choose δ > 0 and free for the moment. Since u+ = v, there is a q > 0 such that ||Uq − v||L∞ [0,1] ≤ δ /2. Hence along our subsequence for all large k, ||uk − U||L∞ [q,q+1] ≤ δ /2. Thus ||v − uk ||L∞ [q,q+1] ≤ δ for large k. Define u∗k to be equal to v for t ≤ q, equal to φk for q ≤ t ≤ q + 1, and equal to uk for q + 1 ≤ t, where φk is a minimizer of the variational problem  q+1

inf

L(u) dt q

378

P.H. Rabinowitz

over K = {u ∈ W 1,2 [q, q + 1] : u(q) = v(q), u(q + 1) = uk (q + 1)}. The minimality properties of v and w imply v ≤ φk ≤ w and therefore u∗k ∈ Γ1 . Set u(t) = v(t) + (t − q)(uk (q + 1) − v(q + 1)) so u ∈ K. Moreover (9) aq (φk ) ≤ aq (u) ≤ β (δ ) where β (δ ) → 0 as δ → 0. Now by (8)–(9) J(u∗k ) − J(uk ) =

∞

∑ [a p (u∗k ) − a p (uk )]

−∞

q

= aq (u∗k ) − ∑ a p (uk ) −∞

≤ β (δ ) −

γ (ε ) . 2

(10)

Choosing δ so small that β (δ ) ≤ 14 γ (ε ), (10) contradicts that {uk } is a minimizing sequence for (2). Thus U ∈ Γ1 , and J(U) ≥ c1 . On the other hand, p

p

∑ ai (U) ≤ lim inf ∑ ai (uk ) ≤ lim inf J(uk ) = c1 k→∞ −p

−p

k→∞

so letting p → ∞, we conclude J(U) = c1 . This establishes statement 1 of Theorem 3.1. To prove statement 2 of Theorem 3.1, first we will obtain the minimality property 5. If it is not true, there are numbers r < s and a function

φ ∈ {u ∈ W 1,2 [r, s] : u(r) = U(r) and u(s) = U(s)}  s

such that

r

L(φ ) dt
U(t + 1), or (iii) U(t) < U(t + 1). If alternative (i) holds, U is 1-periodic and therefore U ∈ / Γ1 while (ii) implies U(t) > U(t + k) → w(t) as k → ∞. Thus U > w and again U ∈ / Γ1 . Thus (iii) holds.

 We conclude this section with a result that shows the gap condition is not only sufficient for there to exist minimizing heteroclinics from v to w, but also is necessary. Theorem 3.2. Let V satisfy (V1 ) - (V2 ), and further let v, w ∈ M0 with v < w. Suppose there is a U ∈ Γ1 (v, w) such that J(U) =

inf

u∈Γ1 (v,w)

J(u).

Then v and w are a gap pair. Proof. Otherwise there is a φ ∈ M0 such that v < φ < w. There is a smallest α ∈ R such φ (α ) = U(α ). Define W (t) = U(t), for t ≤ α , W (t) = φ (t) when α ≤ t ≤ α +1, and W (t) = U(t − 1) when α + 1 ≤ t. Then W ∈ Γ1 (v, w) and J(W ) = J(U). Set S = {u ∈ W 1,2 [α − 1/2, α + 1/2] : u(α ± 1/2) = W (α ± 1/2)}. The remark following Proposition 2.3 shows there is a ψ ∈ S such that ψ is a solution of (DE) and  α +1/2 α −1/2

We claim A≡

L(ψ ) dt = inf

 α +1/2 α −1/2

 α +1/2

u∈S α −1/2

L(ψ ) dt
0. Then there is an m0 (ε ) such that if m2 − m1 , m4 − m3 ≥ m0 (ε ), bm,ρ ≤ c1 (v, w) + c1 (w, v) + ε Proof. Let U¯ ∈ M1 (v, w). Then there are α , β ∈ Z with α ≤ β such that if f¯, g¯ are respectively minimizers of  α

 β +1

α −1

L(u) dt,

β

L(u) dt

over ¯ α )}, {u ∈ W 1,2 [α − 1, α ] : u(α − 1) = v(α − 1), u(α ) = U( ¯ β ), u(β + 1) = w(β + 1)}. {u ∈ W 1,2 [β , β + 1] : u(β ) = U( Then

ε ¯ ≤ . aα −1 ( f¯), aβ (g) 4

(8)

−1 ¯ αβ to g¯ to w|∞ defines U ∗ ∈ Γ1 (v, w) (Fig. 9). Since J(U) ¯ = Gluing v|α−∞ to f¯ to U| β +1 c1 (v, w) by (8),

J(U ∗ ) = aα −1 ( f¯) + ≤ c1 (v, w) +

β −1

¯ + aβ (g) ¯ ∑ ai (U)

ε 2

α

(9)

The calculus of variations and the forced pendulum

383

Fig. 9 The construction of U ∗ in Proposition 8.

Similarly let U ∈ M1 (w, v). As above there are r, s ∈ Z with r < s such that if f , respectively g are the minimizers of  r

 s+1

L(u) dt, r−1

L(u) dt s

over {u ∈ W 1,2 [r − 1, r] : u(r − 1) = w(r − 1), u(r) = U(r)}, {u ∈ W 1,2 [s, s + 1] : u(s) = U(s), u(s + 1) = v(s + 1)}. then

ε ar−1 ( f ), as (g) ≤ . 4

(10)

s ∞ and gluing w|r−1 −∞ to f to U|r to g to v|s+1 produces U∗ ∈ Γ1 (w, v) with

ε J(U∗ ) ≤ c1 (w, v) + . 2

(11)

Finally set U ∗∗ (t) equal to U ∗ (t − m2 + β + 1) for t ≤ m2 , and equal to U∗ (t − m3 + r − 1) for m2 ≤ t. By construction U ∗∗ satisfies the constraints of Ym,ρ at t = m2 and m3 . For m2 − m1 ≥ β − α + 2, U ∗∗ (m1 ) = U ∗ (m1 − m2 + β + 1) = v(α − 1) = v(m1 ) so U ∗∗ satisfies the constraint at t = m1 . Similarly the constraint at t = m4 holds if m4 − m3 ≥ s − r + 2. Therefore U ∗∗ ∈ Ym,ρ and by (9) and (11), bm,ρ ≤ J(U ∗∗ ) ≤ c1 (v, w) + c1 (w, v) + ε



Next we will refine our choice of ρ . Recall M1 (v, w) and M1 (w, v) have gaps. Define ρ− : M1 (v, w) → (0, w(0) − v(0)) via ρ− (u) = u(0) − v(0). Therefore ρ− is a monotone function of u and ρ− (M1 (v, w)) has gaps. Choose ρ1 to lie in such a gap, i.e. ρ1 ∈ (0, w(0) − v(0))\ ρ− (M1 (v, w)). Note that ρ1 can be chosen as small as desired since f ∈ M1 implies f (· − l) ∈ M1 (v, w) for any l ∈ Z so for large l, ρ− ( f (· − l)) is near 0. Similarly define ρ+ : M1 (v, w) → (0, w(0) − v(0)) via ρ+ (u) = w(0) − u(0) so ρ+ is also monotone and ρ+ (M1 (v, w)) has gaps. Choose ρ2 in such a gap. Likewise ρ− , ρ+ : M1 (w, v) → (0, w(0)−v(0)) as above. Choose ρ3 and ρ4 in associated gaps. An important consequence of this choice of the ρi is:

384

P.H. Rabinowitz

Proposition 4.5. Let

Λ1 (v, w) = {u ∈ Γ1 : u(0) − v(0) = ρ1 or w(0) − u(0) = ρ2 }. Set d1 (v, w) =

inf

u∈Λ1 (v,w)

J(u)

(12)

Then for |ρ | small, d1 (v, w) > c1 (v, w). Remark Defining Λ1 (w, v) and d1 (w, v) in the obvious way, we also have d1 (w, v) > c1 (w, v). Proof of Proposition 4.5. Let {un } be a minimizing sequence for (12). As in the 1,2 proof of Theorem 3.1, there is a P ∈ Wloc such that along a subsequence un → P 1,2 ∞ weakly in Wloc and also in Lloc . This latter convergence implies v ≤ P ≤ w and P satisfies one of the constraints at t = 0. Also, as earlier J(P) < ∞ and therefore P asymptotes to v or w as t → −∞ and as t → ∞. If (a) P(0) = v(0) + ρ1 , since ρ1 is small, the argument of Proposition 4.3 shows ||P − v||W 1,2 [i,i+1] → 0 as i → −∞, while if (b) w(0) = P(0) + ρ2 , similarly ||P − w||W 1,2 [i,i+1] → 0 as i → ∞. Suppose (a) holds. Then either (c) ||P − v||W 1,2 [ j, j+1] → 0 as j → ∞ or (d) ||P − w||W 1,2 [ j, j+1] → 0 as j → ∞. If (c) occurs, un (0) is near v(0)+ ρ1 along a subsequence as n → ∞. Hence as in the proof of Lemma 3.1, there is a γ (ρ1 ) > 0 (independent of n) such that (13) a0 (un ) ≥ γ (ρ1 ) for large n. Moreover for any δ > 0, there is an l = l(δ ) ∈ N such that ||un − v||L∞ [l,l+1] ≤ δ for large n along the subsequence. Now in the spirit of the proof of Proposition 4.3, set u∗n equal to v, on t ≤ l, equal to gn on l ≤ t ≤ l + 1, and equal to un on t ≥ l + 1 where gn minimizes  l+1

L(u) dt l

over {u ∈ W 1,2 [l, l + 1] : u(l) = v(l), u(l + 1) = un (l + 1)}. Thus as in (9) again, al (un ) ≤ β (δ ).

(14)

Now by (13)–(14), ∞

J(un ) ≥ a0 (un ) + ∑ ai (un ) l+1 ∞

≥ γ (ρ1 ) + ∑ ai (u∗n ) − al (gn ) l

≥ γ (ρ1 ) + J(u∗n ) − β (δ ).

(15)

The calculus of variations and the forced pendulum

385

Choosing δ so that β (δ ) ≤ 1/2γ (ρ1 ) and noting that u∗n ∈ Γ1 (v, w), (15) yields 1 J(un ) ≥ c1 (v, w) + γ (ρ1 ). 2

(16)

Thus d1 (v, w) ≥ c1 (v, w) + 12 γ (ρ1 ) for this case. On the other hand, if (a) and (d) occur, P ∈ Λ1 (v, w) and by earlier arguments J(P) = d1 (v, w). Since Λ1 (v, w) ⊂ Γ1 (v, w), d1 (v, w) ≥ c1 (v, w). If d1 = c1 , then P ∈ M1 (v, w) and by Theorem 3.1, P is a solution of (DE). Consequently P(0) − v(0) = ρ1 = ρ− (P) ∈ ρ− (M1 (v, w)) contrary to the choice of ρ1 . Thus  d1 > c1 . The remaining cases are treated in the same fashion as above.

With the aid of Proposition 4.5, we have: Proposition 4.6. Set

µ=

1 min(d1 (v, w) − c1 (v, w), d1 (w, v) − c1 (w, v)). 2

If |ρ | is small and U satisfies an mi constraint with equality, then for m3 − m2 >> 1, bm,ρ ≥ c1 (v, w) + c1 (w, v) + µ .

(17)

Assuming Proposition 4.6 for the moment, combining Proposition 4.5 and Proposition 4.6 we have µ ≤ bm,ρ − c1 (v, w) − c1 (w, v) ≤ ε (18) provided that an mi constraint holds with equality. Here µ depends only on ρ and m3 − m2 while m2 − m1 , m4 − m3 ≥ m0 (ε ). Thus choosing ε < µ , (18) yields a contradiction. Therefore U satisfies (DE) for all t and we have Theorem 4.1. If (V1 )–(V2 ) hold, v and w are a gap pair, and in addition M1 (v, w) and M1 (w, v) have gaps, then for |ρ | small and mi+1 − mi large, there is a U = Um,ρ ∈ Ym,ρ which is a solution of (DE) with J(U) = bm,ρ . Remark: That U satisfies the constraints with strict inequality implies U has a local minimization property. Corollary 4.1. There are infinitely many distinct 2-transition solutions of (DE). Proof. Simply take different sets of (mi )’s with mi+1 − mi larger and larger.

 To complete the proof of Theorem 4.1, we give the 3 Proof of Proposition 4.6. By the minimality property of U|m m2 ,

 m3 m2

L(U) dt = inf

 m3

u∈A m2

L(u) dt

where A = {u ∈ W 1,2 [m2 , m3 ] : u(m2 ) = U(m2 ), u(m3 ) = U(m3 )}.

(19)

386

P.H. Rabinowitz

Since ρ2 , ρ3 are small and w(m2 ) − U(m2 ) ≤ ρ2 , w(m3 ) − U(m3 ) ≤ ρ3 , as in (9), (19) implies m3 −1

∑

ai (U) ≤ β (ρ2 ) + β (ρ3 ).

(20)

i=m2

We claim that given any σ > 0, there is an α (σ ) > 0 such that for m3 − m2 ≥ α (σ ), ||U − w||W 1,2 [i,i+1] ≤ σ for some q ∈ [m2 , m3 − 1]. Otherwise by Lemma 3.1, m3 −1

∑

a j (U) ≥ (m3 − m2 − 1)γ (σ )

(21)

j=m2

which goes to infinity as m3 − m2 → ∞. But this is contrary to (20) which shows that the left hand side of (21) is small. Now suppose for convenience that we have equality at an m1 or m2 constraint point. Set Φ (t) equal to U(t) for t ≤ q, equal to f (t) for q ≤ t ≤ q + 1, equal to w(t) for t ≥ q + 1, where f minimizes  q+1

L(u) dt q

over {u ∈ W 1,2 [q, q + 1] : u(q) = U(q), u(q + 1) = w(q + 1)}. Therefore Φ ∈ Γ1 (v, w). Similarly set Ψ (t) equal to w(t) for t ≤ q, equal to g(t) for q ≤ t ≤ q + 1, and equal to U(t) for t ≥ q + 1, where g minimizes  q+1

L(u) dt q

over {u ∈ W 1,2 [q, q + 1] : u(q) = w(q), u(q + 1) = U(q + 1)}. Therefore Φ ∈ Γ1 (w, v) and d1 (v, w) + c1 (w, v) ≤ J(Φ ) + J(Ψ ) ≤ J(U) − aq (U) + aq ( f ) + aq (g)

(22)

Since ||U − w||W 1,2 [q,q+1] ≤ σ it follows as in (9) again that aq ( f ) + aq (g) ≤ 2β (σ ) → 0 as σ → 0. Then for σ so small that 2β (σ ) ≤ µ , J(U) = b and (22)–(23) imply (17).



(23)

The calculus of variations and the forced pendulum

387

5 Multitransition solutions: general case The ideas used in proving Theorem 4.1 work equally well to get k transition solutions of (DE) and then even infinite transition solutions via a limit argument, provided that the construction does not depend on k. However, given Theorem 4.1, there is a simpler geometrical argument giving the k and infinite transition cases, as well as an associated symbolic dynamics of solutions. We will illustrate with the case of k = 3 and then discuss the general case. Choose ρ , r ∈ R4 and m, n ∈ Z4 such that there are associated solutions U and W of (DE) with U ∈ Ym,ρ (v, w), and W ∈ Yn,r (w, v). We seek a 3-transition solution heteroclinic from v to w. For j ∈ Z, set τ j u(t) = u(t − j). Because of their asymptotic properties for j1 >> 1, τ j1 U(t) < W (t) for all t ∈ R. Take j2 >> j1 . Then τ j1 U < τ j2 W . Finally take j3 >> j2 . Then τ j3 U < τ j2 W . For simplicity, we will take j1 = j, j2 = 2 j, j3 = 3 j for sufficiently large j. Consider {τ−l jW : l ∈ N} and {τ(3+i) jU : i ∈ N}. Delete from the region between the graphs of v and w the set of points above all of the shifted W ’s we have mentioned and below the shifted U’s. Denote the remaining region by R and set 1,2 ¯ : (t, u(t)) ∈ R}. Y (R) ≡ {u ∈ Wloc

(See Fig. 10). Define c(R) = inf J(u). u∈Y (R)

Then we have: Theorem 5.1. Under the hypothesis of Theorem 4.1 1. M(R) = {u ∈ Y (R) : J(u) = c(R)} = 0. / 2. Any U ∈ M(R) is a classical solution of (DE) and is interior to R. 3. ||U − v||L2 [i,i+1] → 0, i → −∞, and ||U − w||L2 [i,i+1] → 0, i → ∞.  4. U has a local minimization property: for any r < s, U minimizes rs L(u) dt over the class of W 1,2 [r, s] functions with u(r) = U(r), and u(s) = U(s) provided that s − r sufficiently small. Proof. We will sketch the proof. A minimizing sequence converges as earlier to ¯ < ∞, (3) of the theorem holds due U lying in R¯ with J(U) = c(R). Since J(U) w

U

Fig. 10 A U in Y (R).

v

388

P.H. Rabinowitz

to the form of R. The boundary of R consists of curves possessing local or global minimization properties and this readily implies (4), which in turn gives the first part of (2). Lastly the basic existence and uniqueness theorem for ordinary differential  equations implies U cannot touch ∂ R as in the proof of Theorem 2.2.

Next we will show how to generalize Theorem 5.1 and at the same time get a symbolic dynamics of solutions (Fig. 10). Choose U, W , and j as above so in particular the graphs of τ± jU and W do not intersect. This implies the same is true of the graphs of τi jU and τl jW for all i, l ∈ Z. Define

Σ ≡ {σ = {σi }i∈Z : σi ∈ {+, −}}. For each σ ∈ Σ , we define a region R(σ ) lying between the graphs of v and w as follows. Set S = {(t, z) : t ∈ R, v(t) ≤ z ≤ w(t)}. If σi = +, remove the region below τ jiU from S; if σi = −, remove the region above τ jiW from S. R(σ ) is what remains after carrying out this excision process for all i ∈ Z. Then we have; Theorem 5.2. For each σ ∈ Σ , there is a solution UR(σ ) of (DE) with the graph of UR(σ ) lying in R(σ ). Moreover UR(σ ) has the local minimization property of Theorem 5.2. Remark: If σi = −, UR(σ ) will be L∞ close to v on a large interval while if σi = +, UR(σ ) will be L∞ close to w on a large interval. In particular if σi = − for all i near −∞, UR(σ ) asymptotes to v as t → −∞, while if σi = + for all i near ∞, UR(σ ) asymptotes to w as t → ∞. The dynamics of the symbol σ reflect the dynamics of the solution UR(σ ) . Proof of Theorem 5.2. We will sketch the proof. First we introduce four subsets of Σ: Σ ++ ≡ {σ ∈ Σ : σi = + for all large |i|}

Σ −− ≡ {σ ∈ Σ : σi = − for all large |i|} Σ +− ≡ {σ ∈ Σ : σi = + for all large negative i, and σi = − for all large positive i} Σ −+ ≡ {σ ∈ Σ : σi = − for all large negative i, and σi = + for all large positive i} Let Σ ∗ be the union of these four sets. Any σ ∈ Σ ∗ has a finite number of changes of σi as i increases. For σ ∈ Σ ∗ , set 1,2 ¯ σ ) for all t ∈ R}, : (t, u(t)) ∈ R( Y (σ ) = {u ∈ Wloc

and define c(σ ) = inf J(u). u∈Y (σ )

The calculus of variations and the forced pendulum

389

Then c(σ ) < ∞ and the proof of Theorem 5.1 shows there is a UR(σ ) ∈ Y (σ ) such that UR(σ ) satisfies (2) and (4) of Theorem 5.1 and also possess the asymptotics associated with σ . Next suppose σ = {σi }∈Z ∈ Σ \Σ ∗ . For each n ∈ N, define fn (σ ) ∈ Σ ∗ via fn (σ ) equal to σi , |i| ≤ n, equal to σn , i > n, and equal to σ−n when i < −n. Therefore by what was previously shown, there is a Un ∈ Y ( fn (σ )) such that J(Un ) = c( fn (σ )). Since v ≤ Un ≤ w, the functions Un are uniformly bounded. By (DE), they are also bounded in C2 . Therefore using (DE), as n → ∞, Un converges along a subsequence in C2 to U(σ ), a solution of (DE). Moreover for any l ∈ N, if n ≥ l, for |t| ≤ l, the graph of Un lies in R( fn (σ )) ∩ {(t, z) : |t| ≤ l, v(t) < z < w(t)} = R(σ ) ∩ {(t, z) : |t| ≤ l, v(t) < z < w(t)} from which it follows that the graph of U lies in R(σ ). Finally the local minimality ∞ convergence of the U .

 property is preserved by the Lloc n We conclude this section with some open questions. First, is it possible to give a variational characterization of U(σ ) for σ ∈ Σ \Σ ∗ ? The difficulty is that for such σ , J(U(σ )) = ∞. We suspect that a second renormalization of J can be made which allows for a direct variational characterization of U(σ ). A second question is whether it is possible to classify these multi-transition solutions. How many parameters do they really depend on?

6 The tip of the iceberg In a sense the class of solutions of (DE) we have studied in these lectures merely represent the tip of the iceberg. All of these solutions lie between a gap pair. Even if we had had time to study the monotone solutions of (DE) mentioned in the introduction that cross a finite number of gaps, we are still only dealing with bounded solutions which therefore have rotation number 0. For p ∈ Z and q ∈ N it is straightforward to find minimal solutions of (DE) satisfying u(t + q) = u(t) + p. In terms of the pendulum, they make p rotations in time q and have rotation number p/q. Thus replacing M0 by such a class of minimizers, there are analogues of the results of the previous sections. There are also minimal solutions with an irrational rotation number which can be obtained as limits of the rational ones. In addition to these minimal solutions there are nonminimal solutions that can be obtained variationally. E.g. there are mountain pass solutions lying between a gap pair v, w. In fact there is a sequence {un } of such solutions with periods which go to infinity as n → ∞. Likewise there are mountain pass heteroclinics between a gap pair in M1 (v, w). These facts can be proven using versions of the mountain pass theorem.

390

P.H. Rabinowitz

References [A] [B88] [B89] [Ma82] [Ma93] [Mor] [Mos86] [RS]

S. Aubry, and P.Y. LeDaeron, The discrete Frenkel-Kantorova model and its extensions I-Exact results for the ground states, Physica, 8D (1983), 381–422. V. Bangert, Mather sets for twist maps and geodesics on tori. Dynamics reported, Vol. 1, 1–56 (1988). V. Bangert, On minimal laminations of the torus, AIHP Analyses Nonlinéaire, 6 (1989), 95–138. J.N. Mather, Existence of quasi-periodic orbits for twist homeomorphisms fo the annulu, Topology, 21 (1982), 457–467. J.N. Mather, Variational construction of connecting orbits, Ann. Inst. Fourier (Grenoble) 43 (1993), 1349–1386. H. Marston Morse, A fundamental class of geodesics on any closed surface of genus grater than one. Trans. Amer. Math. Soc. 26 (1924), no. 1, 25–60. J. Moser, Minimal solutions of variational problems on a torus, AIHP Analyse Nonlinéaire, 3 (1986), 229–272. P.H. Rabinowitz, and E. Stredulinsky, Single and multitransition solutions for a class of PDE’s, in progress.

Variational methods for Hamiltonian PDEs Massimiliano Berti1

Abstract We present recent existence results of periodic solutions for completely resonant nonlinear wave equations in which both “small divisor” difficulties and infinite dimensional bifurcation phenomena occur. These results can be seen as generalizations of the classical finite-dimensional resonant center theorems of Weinstein–Moser and Fadell–Rabinowitz. The proofs are based on variational bifurcation theory: after a Lyapunov–Schmidt reduction, the small divisor problem in the range equation is overcome with a Nash–Moser implicit function theorem for a Cantor set of non-resonant parameters. Next, the infinite dimensional bifurcation equation, variational in nature, possesses minimax mountain-pass critical points. The big difficulty is to ensure that they are not in the “Cantor gaps”. This is proved under weak non-degeneracy conditions. Finally, we also discuss the existence of forced vibrations with rational frequency. This problem requires variational methods of a completely different nature, such as constrained minimization and a priori estimates derivable from variational inequalities.

1 Finite dimensions: resonant center theorems Consider a finite dimensional Hamiltonian system x˙ = J∇H(x) ,  where J =

0 I −I 0

x ∈ R2n

(1)

 is the symplectic matrix, I is the identity in Rn , and ∇H(0) = 0.

• Q UESTION: Do there exist periodic solutions of (1) close to the equilibrium x = 0? 1

Dipartimento di Matematica e Applicazioni, R. Caccioppoli, Università Federico II, Via Cintia, Monte S. Angelo, I-80126 Napoli, Italy e-mail: [email protected]

W. Craig (ed.), Hamiltonian Dynamical Systems and Applications, 391–420. c 2008 Springer Science + Business Media B.V.


391

392

M. Berti

Clearly, a necessary condition is the presence of purely imaginary eigenvalues of J(D2 H)(0) so that the linearized equation x˙ = J(D2 H)(0)x

(2)

possesses periodic solutions. In the sequel we suppose that x = 0 is an elliptic equilibrium with J(D2 H)(0) having all eigenvalues ±iω1 , . . . , ±iωn

(3)

purely imaginary. The first continuation result of periodic solutions close to an elliptic equilibrium is the celebrated Lyapunov center theorem. Assuming the non resonance condition ω j − l ω1 = 0, ∀l ∈ Z, j = 2, . . . , n, the theorem ensures the existence of a smooth two-dimensional manifold foliated by small amplitude periodic solutions of (1) with frequencies close to ω1 , see e.g. [32]. If the above non-resonance condition is violated, no periodic solutions except the equilibrium x = 0 need exist. An example due to Moser [31] is provided by the Hamiltonian in (R4 , ∑2i=1 dqi ∧ d pi ) H=

q21 + p21 q22 + p22 − + (q21 + p21 + q22 + p22 )(p1 p2 − q1 q2 ) 2 2

where (q, p) = 0 is an elliptic equilibrium with non-simple eigenvalues ±i. But d (q1 p2 + p1 q2 ) = 4(p1 p2 − q1 q2 )2 + (q21 + p21 + q22 + p22 )2 dt so that the unique periodic solution is q = p = 0. In contrast, two remarkable theorems by Weinstein [42], Moser [31] and Fadell– Rabinowitz [24] prove the existence of periodic solutions under the assumptions respectively (WM)

(D2 H)(0) > 0 ,

(FR)

signature(D2 H)(0) = 0 .

Theorem 1.1. (Weinstein 1973–Moser 1976) Let H ∈ C2 (R2n , R) such that (∇H)(0) = 0 and (D2 H)(0) > 0. For all ε small enough there exist, on each energy surface {H(x) = H(0) + ε 2 }, at least n geometrically distinct periodic solutions of (1). Note that (D2 H)(0) > 0 implies that the level sets of the energy 12 (D2 H)(0)x · x of (2) are ellipsoids, whence all the solutions of (2) are bounded and (3) holds. Theorem 1.2. (Fadell–Rabinowitz 1978) Let H ∈ C2 (R2n , R) such that (∇H)(0) = 0 and the signature(D2 H)(0) = 2ν = 0. Assume also that any non-zero solution of (2) is T -periodic (and not constant). Then, either (i) x = 0 is a non-isolated T -periodic solution of (1), or

Variational methods for Hamiltonian PDEs

393

(ii) there exist integers k, m ≥ 0 with k + m ≥ |ν | and a left neighborhood, Ul , resp. a right neighborhood, Ur , of T in R such that ∀λ ∈ Ul \{T }, resp. Ur \{T }, there exist at least k, resp. m, distinct, non-trivial, λ -periodic solutions of (1). The L∞ -norm of the solutions tends to 0 as λ → T . Unlike the Lyapunov Center Theorem the periodic solutions of Theorems 1.1 and 1.2 do not, in general, vary smoothly with respect to the parameters ε (energy) and λ (period). This will cause a serious difficulty for PDEs, see Sections 5 and 6. Both the Weinstein–Moser and the Fadell–Rabinowitz resonant center theorems follow by arguments of variational bifurcation theory. The Weinstein–Moser theorem, thanks to the energy constraint, reduces to find critical points of a smooth function defined on a (2n − 1)-dimensional sphere. The Fadell–Rabinowitz theorem is more subtle because there is no energy constraint, and one has to look for critical points of a reduced action functional defined in an open neighborhood of the origin, see e.g. [4]. For brevity, we present only a simplified version of the Fadell–Rabinowitz theorem without obtaining the optimal multiplicity results. We shall also assume that (D2 H)(0) := I

(4)

so that ±i are the multiple eigenvalues of J(D2 H)(0), and (2) possesses a 2ndimensional linear space of periodic solutions with the same minimal period 2π . Theorem 1.3. (Fadell–Rabinowitz) Under the assumptions above, either (i) x = 0 is a non-isolated 2π -periodic solution of (1), (ii) There is a one sided neighborhood U of 1 such that, ∀λ ∈ U \ {1}, equation (1) possesses at least two distinct non-trivial 2πλ -periodic solutions, (iii) There is a neighborhood U of 1 such that, ∀λ ∈ U \ {1}, equation (1) possesses at least one non-trivial 2πλ -periodic solution. For the extension of these results to PDEs, due to the necessity of imposing nonresonance conditions on the frequency, it will be more convenient to give existence results with fixed frequency like in the Fadell–Rabinowitz theorem (see Theorems 3.2, 5.1, 6.1), and not with fixed energy as in the Weinstein–Moser theorem.

1.1 The variational Lyapunov–Schmidt reduction With no loss of generality we suppose that H(0) = 0. Normalizing the period we look for 2π -periodic solutions of J x˙ + λ (∇H)(x) = 0 .

(5)

394

M. Berti

Equation (5) is the Euler–Lagrange equation of the action functional

Ψ (λ , x) :=


1 T

2

 J x(t) ˙ · x(t) + λ H(x(t)) dt ,

T := (R/2π Z)

defined and C1 , e.g. on the Sobolev space H 1 (T). To find critical points of Ψ we perform a Lyapunov–Schmidt reduction, decomposing

where V := v ∈ H 1 (T) | v˙ = J(D2 H)(0)v H 1 (T) := V ⊕V ⊥ 

is 2n-dimensional (by (4)) and V ⊥ := {w ∈ H 1 (T) | T w · v dt = 0 , ∀v ∈ V }. Projecting (5), for x = v + w, v ∈ V , w ∈ V ⊥ , yields & ΠV (J(v˙ + w) ˙ + λ (∇H)(v + w)) = 0 bifurcation equation ΠV ⊥ (J(v˙ + w) ˙ + λ (∇H)(v + w)) = 0 range equation where ΠV , ΠV ⊥ denote the projectors on V , resp. V ⊥ . The range equation. We solve first the range equation with the standard implicit function theorem, finding a solution w(λ , v) ∈ V ⊥ for v ∈ Br (0) (≡ ball in V of radius r centered at zero), r > 0 small enough, and λ sufficiently close to 1. Indeed ˙ + λ ∇H(v + w)) F (λ , v, w) := ΠV ⊥ (J(v˙ + w) vanishes F (λ , 0, 0) = 0, ∀λ , and its partial derivative (Dw F )(1, 0, 0)[W ] = JW˙ + (D2 H)(0)W ,

∀W ∈ V ⊥

is an isomorphism. By the implicit function theorem the solution w(λ , v) ∈ V ⊥ of the range equation is a C1 function, w(λ , 0) = 0 and w(λ , v) = o(v)

as v → 0

(6)

uniformly for λ near 1. The bifurcation equation. It remains to solve the bifurcation equation

ΠV (J(v˙ + w( ˙ λ , v)) + λ (∇H)(v + w(λ , v))) = 0 which is the Euler–Lagrange equation of the “reduced action functional”

Φ (λ , ·) : Br (0) ⊂ V → R ,

Φ (λ , v) := Ψ (λ , v + w(λ , v)) .

Indeed ∀h ∈ V , (Dv Φ )(λ , v)[h] = (DxΨ )(λ , v + w(λ , v))[h + (Dv w)(λ , v)[h] ] = (DxΨ )(λ , v + w(λ , v))[h] 

= T

ΠV (J(v˙ + w) ˙ + λ ∇H(v + w)) · h dt

(7)

Variational methods for Hamiltonian PDEs

using that (DxΨ )(λ , v + w(λ , v))[W ] = 0 ,

395

∀W ∈ V ⊥

since w(λ , v) solves the range equation, and (Dv w)(λ , v)[h] ∈ V ⊥ . Remark also that, by (7) and since w(λ , v) ∈ C1 , then Dv Φ ∈ C1 and so Φ ∈ C2 . To prove Theorem 1.3 we have to find non-trivial critical points of the reduced action functional Φ (λ , ·) near v = 0 for fixed λ near 1. The functional Φ (λ , ·) possesses a strict local minimum or maximum at v = 0, according λ > 1 or λ < 1, because, by (6), and H(0) = (∇H)(0) = 0, (D2 H)(0) = I, 

1 λ J v(t) ˙ · v(t) + (D2 H)(0)v(t) · v(t) + o(v2 ) 2 2  (λ − 1) |v(t)|2 dt + o(v2 ) = (λ − 1)π |v(0)|2 + o(v2 ) . = 2 T

Φ (λ , v) =

T

(8)

If v = 0 is not an isolated critical point of Φ (1, ·) then alternative (i) of Theorem 1.3 holds. If v = 0 is an isolated critical point of Φ (1, ·) then, either (a) Φ (1, ·) has a strict local maximum or minimum in v = 0 (b) Φ (1, ·) takes on both positive and negative values near v = 0. Case (a) leads to alternative (ii) and Case (b) leads to alternative (iii) of Theorem 1.3. Figure 1 gives the idea of the existence proof. In both cases (a) and (b), the functional Φ (λ , ·) possesses saddle critical points which can be found by a (finite dimensional) mountain pass argument [1] (see Theorem 3.1). However, the main difficulty of the minimax proof is that Φ (λ , ·) is defined only in a neighborhood of zero, see [4], [37] for details. In the following we shall develop an analogous variational argument in infinite dimension.

Fig. 1 In case (a), Φ (1, ·) has a strict local maximum at v = 0 and for λ > 1, Φ (λ , ·) possesses at least two non-trivial critical points. In case (b) Φ (1, ·) takes on both positive and negative values near v = 0 and for λ = 1, Φ (λ , ·) possesses at least one mountain pass critical point.

396

M. Berti

2 Infinite dimensions We want to extend the local bifurcation theory of periodic solutions described in the previous section to infinite dimensional Hamiltonian PDEs, like the “completely resonant” nonlinear wave equation utt − uxx = f (x, u),,

u(t, 0) = u(t, π ) = 0

(1)

where f (x, 0) = (∂u f )(x, 0) = 0. All the solutions of the linearized equation at u = 0 utt − uxx = 0,,

u(t, 0) = u(t, π ) = 0

(2)

are 2π -periodic. They can be represented like (Fourier method) v(t, x) =

∑ a j cos( jt + θ j ) sin( jx) ,

aj ∈ R

j≥1

or like superposition of waves traveling in opposite directions (D’Alembert method) v(t, x) = η (t + x) − η (t − x) where η is any 2π -periodic function. This is the infinite dimensional analogous situation considered by the Weinstein–Moser and Fadell–Rabinowitz resonant Center theorems. In trying to extend the Lyapunov–Schmidt reduction scheme of the previous section, we encounter two new problems: • A “small divisor” problem in the range equation • The presence of an infinite dimensional bifurcation equation The “small divisor” problem is that the eigenvalues of ∂tt − ∂xx in a space of functions 2π /ω -periodic in time and satisfying Dirichlet boundary conditions, are −ω 2 l 2 + j 2 ,

l ∈ Z, j ≥ 1.

(3)

Therefore, for almost every ω ∈ R, such eigenvalues accumulate to zero, the inverse (∂tt − ∂xx )−1 is unbounded and the standard implicit function theorem fails. Remark 2.1. When ω ∈ Q the spectrum is not dense in R. We shall discuss this case in Section 7. Existence of periodic solutions of wave equations with a rational frequency ω ∈ Q has been proved via global minimax methods in [36] and [15]. / Q these proofs fail for a lack of compactness introduced by the small When ω ∈ divisors. For some results in the irrational case, see [22]. The small divisors problem for Hamiltonian PDEs was first solved by Kuksin [28] and Wayne [41] using KAM theory and by Craig–Wayne [20] – who were the first to introduce the Lyapunov–Schmidt reduction method for PDEs – via a Nash– Moser implicit function technique. Other existence results of quasi-periodic solutions have been obtained by Bourgain [11, 12, 14] extending the Craig–Wayne [20]

Variational methods for Hamiltonian PDEs

397

approach, and, via KAM theory, e.g. in Kuksin–Pöeschel [18], Eliasson–Kuksin [23] and Yuan [45]. See also the books [18], [29] and references therein. The first existence result for completely resonant wave equations with Dirichlet boundary conditions was obtained in [3] for f (u) = u3 (under periodic boundary conditions we quote [27]). The small divisors problem is bypassed for the zero measure set Wγ of the frequencies defined in (5) of section 3. The choice of f (u) = u3 is required by a non-degeneracy condition to solve the bifurcation equation. In [6], existence of periodic solutions, for the same zero measure set of frequencies Wγ , but for a general nonlinearity, was proved. The key to remove the non-degeneracy condition is to solve the infinite dimensional bifurcation equation via variational methods. We now present these results for f (x, u) := u p . This situation (where no small divisors appear) is a preparation to understand the more difficult case considered in Section 6 where we shall find solutions of the bifurcation equation on a Cantor set using variational methods, for positive measure sets of frequencies.

3 The variational Lyapunov–Schmidt reduction Normalizing the period, and rescaling the amplitude u → δ u, we look for 2π periodic solutions of

ω 2 utt − uxx = ε u p ,

u(t, 0) = u(t, π ) = 0,,

(1)

where ε := δ p−1 , p ≥ 2, in the Banach algebra

X := u ∈ H 1 (Ω , R) ∩ L∞ (Ω , R) | u(t, 0) = u(t, π ) = 0, u(−t, x) = u(t, x) where Ω := T × [0, π ], endowed with norm u := u∞ + uH 1 . We consider here the easier case when p is odd. Equation (1) is the Euler–Lagrange equation of the Lagrangian action functional Ψ ∈ C1 (X, R) defined by

Ψ (u) :=

 2π

dt 0

 π! 2 " ω 2 1 2 ut − ux + ε F(u) dx 0

2

2

where F(u) := u p+1 /p + 1, sometimes called the “Percival variational principle”. / Q this functional is highly non compact (see Remark 2.1). Therefore, to When ω ∈ find its critical points we perform a Lyapunov–Schmidt reduction, decomposing X = V ⊕W where



 V := v = η (t + x) − η (t − x)  η (·) ∈ H 1 (T), η odd

(2)

398

M. Berti

are the solutions in X of the linear equation (2) of section 2 and



W := w ∈ X | (w, v)L2 = 0, ∀ v ∈ V = ∑ wl, j cos(lt) sin jx ∈ X . l≥0, j≥1,l = j

Remark 3.1. Recalling (2) and the compact embedding H 1 (T) → L∞ (T), also the embedding (V,  · H 1 ) → (V,  · ∞ ) is compact. Projecting (1), for u := v + w with v ∈ V , w ∈ W , yields

ω 2 vtt − vxx = εΠV f (v + w) ω 2 wtt − wxx = εΠW f (v + w)

bifurcation equation range equation

(3) (4)

where ΠV : X → V , ΠW : X → W are the projectors respectively on V , W . The range equation. We first solve the range equation assuming that 

γ  ω ∈ Wγ := ω ∈ R  |ω l − j| ≥ , ∀(l, j) ∈ N × N , j = l . l

(5)

Lemma 3.1. [3], [22] For 0 < γ ≤ 1/4 the set Wγ is uncountable, has zero measure and accumulates to ω = 1 both from the left and from the right. Remark 3.2. By the Dirichlet Theorem, there are no real numbers ω such that |ω l − j| ≥ (γ /l τ ), ∀l = j, if τ < 1. For ω ∈ Wγ the eigenvalues of Lω := ω 2 ∂tt − ∂xx restricted to W , satisfy

γ | − ω 2 l 2 + j2 | = |ω l − j||ω l + j| ≥ |ω l + j| ≥ γω , l

∀l = j

whence the inverse −1 w := Lω

∑

j≥1,l≥0,l = j

wl, j 2 −ω l 2 + j2

cos(lt) sin( jx) ,

∀w ∈ W

−1 w ≤ Cγ −1 w. Fixed points of the nonlinear operator is a bounded operator, Lω

G :W →W ,

−1 ΠW f (v + w) G (ε , ω ; w) := ε Lω

are solutions of the range equation. By the contraction mapping theorem we get Lemma 3.2. (Solution of the range equation) Assume ω ∈ Wγ ∩ (1/2, 3/2). ∀R > 0, ∃ ε0 (R) > 0, C0 (R) > 0 such that ∀v ∈ B2R := {v ∈ V | vH 1 ≤ 2R}, ∀0 ≤ εγ −1 ≤ ε0 (R) there exists a unique solution w(ε , v) ∈ W of the range equation satisfying w(ε , v) ≤ C0 (R)εγ −1 . Moreover the map v → w(ε , v) is in C1 (B2R ,W ).

Variational methods for Hamiltonian PDEs

399

3.1 The bifurcation equation It remains to solve the infinite dimensional bifurcation equation

ω 2 vtt − vxx = εΠV f (v + w(ε , v)) which, arguing as in (7) of section 1.1 (see also Figure 2), is the Euler–Lagrange equation of the reduced Lagrangian action functional Φε ∈ C2 (B2R , R) defined by

Φε (v) := Ψ (v + w(ε , v)) . To find critical points of Φε we expand

because





1 ω2 (vt + (w(ε , v))t )2 − (vx + (w(ε , v))x )2 + ε F(v + w(ε , v)) 2 Ω 2  ω 2 2 v2x ε vt − + ε F(v + w(ε , v)) − f (v + w(ε , v))w(ε , v) = 2 2 Ω 2

Φε (v) =

Ω vt wt

=



Ω vx wx



Ω

= 0 and, since w(ε , v) solves the range equation,

ω 2 wt2 − w2x + ε f (v + w(ε , v))w(ε , v) = 0 .

Hence, using that vt 2L2 = vx 2L2 = v2H 1 /2,

ω2 − 1 v2H 1 + ε Φε (v) = 4

 ! Ω

F(v + w(ε , v)) −

" 1 f (v + w(ε , v))w(ε , v) . 2

(6)

Imposing the “frequency-amplitude” relation ω 2 − 1 = −2ε , we get

Φε (v) = −ε

!1 2

v2H 1 −

 Ω

" v p+1 + Rε (v) p+1

ε

Fig. 2 Since w(ε , v) solves the range equation then v + w(ε , v) is a critical point of the functional W  w → Ψ (v + w), i.e. the gradient (∇Ψ )(v + w(ε , v)) is parallel to V .

400

M. Berti ε

R

1

Fig. 3 The mountain Pass geometry of Φε .

where, for some constant C1 (R) > 0,

ε |Rε (v)| , |(∇Rε (v), v)| ≤ C1 (R) . γ

(7)

The functional (that we still denote by Φε ) 1 Φε (v) := v2H 1 − 2

 Ω

v p+1 + Rε (v) p+1

(8)

possesses a local minimum at the origin and one could think to prove existence of non-trivial critical points via the Mountain Pass Theorem 3.1 below, see Figure 3. Note that, since p is odd, Ω v p+1 > 0, ∀v = 0. Theorem 3.1. (Mountain Pass [1]) Let (X,  · ) be a Banach space. Suppose Φ ∈ C1 (X, R) and (i) Φ (0) = 0 (ii) ∃ ρ , α > 0 such that Φ (x) ≥ α if x = ρ (iii) ∃ v ∈ X with v > ρ such that Φ (v) < 0 Define the “Mountain Pass” value c := infγ ∈Γ maxt∈[0,1] Φ (γ (t)) ≥ α where Γ is the minimax class Γ := {γ ∈ C([0, 1], X) | γ (0) = 0 , γ (1) = v}. Then there exists a Palais–Smale sequence xn of Φ at the level c, i.e. Φ (xn ) → c, ¯ = 0 and Φ (x) ¯ = c. DΦ (xn ) → 0. If, up to subsequence, xn → x¯ then DΦ (x) Theorem 3.1 cannot be directly applied because Φε is defined only close to 0. #ε ∈ C2 (V, R) as Step 1: Extension of Φε . We define the extended functional Φ #ε (v) := Φ

v2H 1  v p+1 − + R#ε (v) 2 Ω p+1

where R#ε (v) := λ (v2H 1 R−2 )Rε (v) and λ : [0, +∞) → [0, 1] is a smooth cut-off function with λ (x) = 1 if |x| ≤ 1, λ (x) = 0 if |x| ≥ 4, and |λ (x)| < 1. By definition ⎧ ⎪ for vH 1 ≤ R ⎨ Φε (v) #ε (v) = v2 1  v p+1 (9) Φ H ⎪ for vH 1 ≥ 2R . ⎩ 2 − Ω p+1

Variational methods for Hamiltonian PDEs

401

Furthermore by (7) and the definition of λ

ε |R#ε (v)| , |(∇R#ε (v), v)| ≤ C1 (R) . γ

(10)

#ε has the Mountain Pass geometry. Let 0 < ρ < R. For all v 1 = ρ Step 2: Φ H (8) #ε (v) (9) Φ = Φε (v) =

(7) 1 v2H 1  v p+1 ε − + Rε (v) ≥ ρ 2 − κ1 ρ p+1 − C1 (R) . 2 2 γ Ω p+1

Fix ρ > 0 such that (ρ 2 /2) − κ1 ρ p+1 ≥ ρ 2 /4. For 0 < εγ −1C1 (R) ≤ ρ 2 /8 #ε (v) ≥ 1 ρ 2 > 0 if v 1 = ρ Φ H 8 verifying the assumption (ii) of Theorem 3.1. Let us verify assumption (iii). By (9) t large enough such that for every v0 H 1 = 1, there exists #  # t p+1 t2 # #ε (# Φ t v0 ) = − v0 p+1 < 0 2 p+1 Ω

because p ≥ 3 is an odd integer. Define v# := # t v0 and the Mountain Pass level #ε (γ (s)) > 0 cε := inf max Φ γ ∈Γ s∈[0,1]

(11)

where Γ := {γ ∈ C([0, 1],V ) | γ (0) = 0 , γ (1) = v#} . By the Mountain Pass Theorem #ε at the level cε > 0, 3.1 there is a Palais–Smale sequence vn ∈ V for Φ #ε (vn ) → 0 . #ε (vn ) → cε , ∇Φ Φ

(12)

Step 3: Confinement of the Palais–Smale sequence. By (11) and (10)  ! s2 " s p+1 #ε (s# v#p+1 + 1 =: κ # v2H 1 − v) ≤ max cε ≤ max Φ p+1 Ω s∈[0,1] s∈[0,1] 2

(13)

for 0 < C1 (R)γ −1 ε < 1. Then
 # (∇R#ε (vn ), vn ) #ε (vn ) − (∇Φε (vn ), vn ) = 1 − 1 vn 2H 1 + R#ε (vn ) − Φ p+1 2 p+1 p+1  (10)
1 1 − vn 2H 1 − 1 ≥ (14) 2 p+1 for 0 < 2C1 (R)εγ −1 ≤ 1. By (12), for n large, (13) # #ε (vn ) − (∇Φε (vn ), vn ) ≤ (cε + 1) + vn  1 ≤ κ + 1 + vn  1 Φ H H p+1

402

M. Berti

1 and, by (14), we derive κ + 1 + vn H 1 ≥ ( 12 − p+1 )vn 2H 1 − 1. Hence, there exist R∗ > 0 independent of ε and C∗ (R) > 0 such that

vn H 1 ≤ R∗

for

0 < εγ −1 < C∗ (R) .

(15)

Step 4: Existence of a nontrivial critical point. Fix R¯ := R∗ + 1 and take 0 < εγ −1 ≤ #ε (vn ) = Φε (vn ). Hence ¯ By (15), definitively for n large, vn ∈ BR¯ , and so Φ C∗ (R). #ε (vn ) = ∇Φε (vn ) = vn − ∇G(vn ) + ∇Rε (vn ) → 0 ∇Φ 

1 p+1 . where we have set G(v) := p+1 Ωv By the compact embedding (V,  · H 1 ) → (V,  · ∞ ) (see Remark 3.1) we easily deduce that ∇G : V → V and ∇Rε : BR¯ → V are compact operators. Therefore, since the Palais–Smale sequence vn is bounded in H 1 , vn is precompact and converges to a nontrivial critical point vε of Φε . We have finally proved

Theorem 3.2. [6] Let f (u) = u p for an odd integer p ≥ 3. ∀ω ∈ Wγ with |ω − 1|γ −1 small enough, and ω < 1, equation (1) of section 2 possesses at least one, non trivial, 2π /ω -periodic, small amplitude periodic solution uω . Multiplicity of solutions can also be proved. To deal with nonlinearities f (u) = u p with p even integer is more difficult because  Ω

v p+1 ≡ 0 ,

∀v ∈ V

(16)

(see Lemma 7.3) and so the development (8) no longer implies the Mountain Pass geometry of Φε . To find critical points of Φε we have to develop at higher orders in ε the non-quadratic term, see [6]. This greater difficulty for finding periodic solutions when the nonlinearity f is non-monotonic (under Dirichlet boundary conditions) is a common feature for nonlinear wave equations, see e.g. [34, 36], [15, 16], and Section 7. In physical terms there is no “confinement effect” due to the potential. Actually, in [10] a nonexistence result is proved for even power nonlinearities in the case of spatial periodic boundary conditions, highlighting that the existence result in case of Dirichlet conditions, is due to a “boundary effect”. We also remark that, since V is infinite dimensional, a Fadell–Rabinowitz type argument, working for any nonlinearity, does not apply.

4 The small divisor problem To prove existence of periodic solutions of & ω 2 utt − uxx = ε a(x)u p u(t, 0) = u(t, π ) = 0

(1)

Variational methods for Hamiltonian PDEs

403

for positive measure set of frequencies, we have to relax the non-resonance condition in (5) of section 3 requiring |ω l − j| ≥

γ , lτ

∀(l, j) ∈ N × N , j = l

for some τ > 1. However, for such ω we get just the bound |ω 2 l 2 − j2 | = |ω l − j|(ω l + j) ≥

γ γω (ω l + j) ≥ τ −1 . lτ l

As a consequence, (Lω )−1 |W “loses (τ − 1)-derivatives” and the standard implicit function theorem fails. This small divisor problem is overcome by a Nash–Moser iteration scheme. We look for 2π -periodic solutions of (1) in the Banach algebra (for s > 1/2) Xσ ,s := u(t, x) = ∑ eilt ul (x) | ul ∈ H01 ((0, π ), R) , ul (x) = u−l (x) l∈Z

and

u2σ ,s

:=

∑ e2σ |l| (l 2s + 1)ul 2H 1 < +∞

.

l∈Z

For σ > 0, s ≥ 0, Xσ ,s is the space of all even, 2π -periodic in time functions with values in H01 ((0, π ), R), which have a bounded analytic extension in the complex strip |Im t| < σ with trace function on |Im t| = σ belonging to H s (T, H01 ((0, π ), C)), see [33] (recently in [9] we proved, when the nonlinearity is just Ck w.r.t. u, also existence of solutions for σ = 0, i.e. just Sobolev in time). Projecting (1) according to the orthogonal decomposition Xσ ,s = (V ∩ Xσ ,s ) ⊕ (W ∩ Xσ ,s ) and imposing the “frequency-amplitude” relation ω 2 − 1 = 2s∗ ε with s∗ = ±1 to be chosen later, yields & ∆ v = s∗ ΠV f (v + w) bifurcation equation (2) range equation Lω w = εΠW f (v + w) where ∆ v := vxx + vtt , f (u) := a(x)u p and Lω := ω 2 ∂tt − ∂xx . When ε = 0 we get the “0th-order bifurcation equation”

∆ v = s∗ ΠV (a(x)v p )

(3)

which is the Euler–Lagrange equation of the functional Φ0 : V → R

Φ0 (v) :=

 v2H 1 v p+1 + s∗ a(x) , 2 p+1 Ω

Ω := T × [0, π ] .

404

M. Berti

For definiteness we suppose there exists v˜ ∈ V such that

 Ω

a(x)v˜ p+1 < 0 and so

t > 0 be large enough such that we choose s∗ = 1 (otherwise we take s∗ = −1). Let # Φ0 (# t v) ˜ < 0. The mountain pass value c := inf

max Φ0 (γ (s)) | γ ∈ C([0, 1],V ) , γ (0) = 0 , γ (1) = # t v˜

s∈[0,1]

is a critical level with a non-trivial mountain pass critical set

Kc := v ∈ V | Φ0 (v) = c , ∇Φ0 (v) = 0

(4)

which is compact for the H 1 -topology (because, with the same arguments used in Section 3.1, any Palais–Smale sequence of Φ0 is precompact). In particular Kc is bounded: there exists Rc > 0 such that vH 1 ≤ Rc ,

∀v ∈ Kc .

Remark 4.1. Actually Φ0 has an unbounded sequence of critical levels tending to plus infinity [1], giving rise to multiplicity of periodic solutions of (1) close to the corresponding critical sets of Φ0 . Since V is infinite dimensional a serious difficulty arises. If v ∈ V ∩ Xσ ,s then the solution w(ε , v) of the range equation, obtained with any Nash–Moser iteration scheme will have a lower regularity, e.g. w(ε , v) ∈ Xσ /2,s . Therefore, in solving next the bifurcation equation substituting w = w(ε , v), the best estimate we can obtain is v ∈ V ∩ Xσ /2,s+2 which makes the scheme incoherent. We overcome this difficulty thanks to a reduction onto a finite dimensional bifurcation equation on a subspace of V of dimension N independent of ω decomposing V = V1 ⊕V2 where

⎧ ⎨ V1 := v ∈ V | v(t, x) = ∑Nl=1 cos(lt)ul sin(lx)

⎩ V2 := v ∈ V | v(t, x) = ∑ l>N cos(lt)ul sin(lx)

“low Fourier modes” “high Fourier modes” . (5)

Setting v := v1 + v2 , v1 ∈ V1 , v2 ∈ V2 , system (2) (with ⎧ ⎪ ∆ v = ΠV1 f (v1 + v2 + w) ⎪ ⎨ 1 ∆ v2 = ΠV2 f (v1 + v2 + w) ⎪ ⎪ ⎩ Lω w = εΠW f (v1 + v2 + w)

s∗

= 1) is equivalent to

(Q1) (Q2) range equation

where ΠVi : Xσ ,s → Vi (i = 1, 2) denote the projectors on Vi .

(6)

Variational methods for Hamiltonian PDEs

405

The (Q2)-equation. We find first the solution v2 (v1 , w) of the (Q2)-equation as a fixed point of v2 = ∆ −1 ΠV2 f (v1 + v2 + w) by a contraction mapping argument, thanks to the compactness of the operator ∆ −1 . ¯ σ¯ > 0, such that Lemma 4.1. (Solution of the (Q2)-equation) There exist N, ∀σ ∈ [0, σ¯ ], ∀v1 H 1 ≤ 2Rc , ∀wσ ,s ≤ Rc , there exists a unique solution v2 (v1 , w) ∈ Xσ ,s+2 of the (Q2)-equation with v2 (v1 , w)σ ,s ≤ Rc /2. The function v2 (·, ·) is C∞ and ∀v ∈ Kc , ΠV2 v = v2 (ΠV1 v, 0). Intuitively, to find solutions of the complete bifurcation equation close to the solutions Kc of the 0th order bifurcation equation (3), N¯ must be taken large enough so that the majority of the “H 1 -mass” of the functions in Kc is “concentrated” on the first N¯ Fourier modes. In the sequel we consider as fixed the constants N¯ and σ¯ which depend only on a(x)u p and Kc . The range equation. We solve next the range equation Lω w = εΠW Γ (v1 , w)

(7)

where Γ (v1 , w) := f (v1 + w + v2 (v1 , w)) via a Nash–Moser implicit function theorem. We set B(2Rc ;V1 ) := {v1 ∈ V1 | v1 H 1 ≤ 2Rc }. Theorem 4.1. (Solution of the range-equation) For ε0 > 0 small enough, there exists # ·) ∈ C∞ ([0, ε0 ] × B(2Rc ;V1 ),W ∩ Xσ¯ /2,s ) w(·, # ε , v1 )σ¯ /2,s ≤ Cεγ −1 and the “large” Cantor set B∞ ⊂ [0, ε0 ] × such that w( # ε , v1 ) solves the range equation B(2Rc ;V1 ) defined below, such that ∀(ε , v1 ) ∈ B∞ , w( (7). The Cantor set B∞ is written explicitly as     B∞ := (ε , v1 ) ∈ [0, ε0 ] × B(2Rc ;V1 ) : ω l − j ≥

2γ , (l + j)τ

(8)



# ε , v1 ))  M(v1 , w( 2γ 1  , ∀ j ≥ 1 , ∀l ≥ , l = j ω l − j + ε ≥ 2j (l + j)τ 3ε √ where ω = 1 + 2ε and M(v1 , w) :=

1 2π 2

 Ω

(∂u f )(x, v1 + w + v2 (v1 , w)) dxdt .

To understand how the Cantor set B∞ arises, we recall that the core of any Nash– Moser convergence method (based on a Newton’s iteration scheme) is the proof of the invertibility of the linearized operators h → L (ε , v1 , w)[h] := Lω h − εΠW DwΓ (v1 , w)[h]

406

M. Berti

where w is the approximate solution obtained at a given stage of the Nash–Moser iteration. The operators L (ε , v1 , w) are self-adjoint. Their eigenvalues can be estimated by
ε  . λl j (ε , v1 ) = −ω 2 l 2 + j2 − ε M(v1 , w) + O j The linear operator L (ε , v1 , w) shall be invertible only for the (ε , v1 ) where all the λl j (ε , v1 ) = 0. This is the phenomenon giving rise to the Cantor set of “nonresonant” parameters B∞ . Some further work has to be done to get estimates for the inverse operators in  σ ,s norms. Our approach [7, 9] is different than in [20] and works also for not odd nonlinearities f with low regularity, unlike [20] works for nonlinearities which are odd and analytic in (x, u). We underline that w( ˜ ε , v1 ) is defined for all the (ε , v1 ) ∈ [0, ε0 ] × B(2Rc ,V1 ) and ˜ ·) is a Whitney smooth interpolanot only on the Cantor set B∞ . The function w(·, tion. The “Cantor gaps” in B∞ are the main new problem to solve the bifurcation equation.

5 The (Q1)-equation The last step is to find solutions of the finite dimensional (Q1)-equation

∆ v1 = ΠV1 G (ε , v1 )

(1)

# ε , v1 ) + v2 (v1 , w( # ε , v1 ))) such that (ε , v1 ) belong to the where G (ε , v1 ) := f (v1 + w( Cantor set B∞ . Critical points of the C∞ “reduced Lagrangian action functional”
 # : B(2Rc ;V1 ) → R , Φ # (ε , v1 ) := Ψ v1 + v2 (v1 , w( # ε , v1 )) + w( # ε , v1 ) Φ such that (ε , v1 ) ∈ B∞ are solutions of (1) (“Percival” reduced variational principle). # (ε , ·) possesses, for any ε small enough, a mountain pass As in Section 3.1, Φ critical point v1 (ε ) close to ΠV1 Kc , where Kc is defined in (4) of section 4, as the mountain pass critical set of Φ0 . However, if Kc does not reduce to a non-degenerate solution of (3), then v1 (ε ) could vary in a highly irregular way as ε → 0, the only information available in general is that v1 (ε ) → ΠV1 Kc as ε → 0. Therefore for each ε the mountain pass critical point v1 (ε ) could belong to the complement of the Cantor set B∞ in which the range equation (7) has been solved. The section Eε := {v√1 | (ε , v1 ) ∈ B∞ } ≡ B(2Rc ,V1 ) has “no gaps”, if and only if the frequency ω (ε ) = 1 + 2ε belongs to the zero-measure set Wγ defined in (5) of section 3. This is why in section 3 we have proved the existence result for any nonlinearity f (u) = u p . It can be shown that in between two sections Eε1 , Eε2 such that ω (ε1 ), ω (ε2 ) ∈ Wγ , the complement of B∞ is arcwise connected. Therefore it would be not sufficient

Variational methods for Hamiltonian PDEs

407

to find just a continuous path of solutions ε → v1 (ε ) of equation (1), to conclude the existence of solutions for a positive measure set of frequencies. This is the common principal difficulty in applying variational methods in a problem with small divisors. The Arnold non-degeneracy condition. The simplest situation occurs when at least one solution v¯ ∈ V of (3) of section 4 is non degenerate, i.e.

(v) ¯ = {0} . ker Φ0|V

(2)

This condition is somehow analogous to the “Arnold non-degeneracy condition” in KAM theory, see e.g. [39] (in [20, 21] it is called “twist condition” or condition of “genuine nonlinearity”). For ε = 0 the (Q1)-equation (1) reduces to the projection of equation (3) of section 4 on V1 , namely
 (3) ∆ v1 = ΠV1 a(x)(v1 + v2 (v1 , 0)) p . By the Arnold condition (and Lemma 4.1), v¯1 := ΠV1 v¯ is a non-degenerate solution of (3) and, applying the implicit function theorem, there exists a C∞ path ε → v1 (ε ) of solutions of (1) with v1 (0) = v¯1 . For all ε belonging to the Cantor-like set

(4) C := ε ∈ [0, ε0 ) | (ε , v1 (ε )) ∈ B∞ they give rise to solutions of equation (1) of section 4 like # ε , v1 (ε ))) + w( # ε , v1 (ε )) ∈ Xσ¯ /2,s . u(ε ) = v1 (ε ) + v2 (v1 (ε ), w( By the smoothness of v1 (·), the set C has asymptotically full measure, namely lim

η →0+

|C ∩ [0, η )| = 1. η

(5)

Geometrically this estimate exploits the structure of the Cantor set B∞ and that the curve of solutions ε → v1 (ε ) crosses transversally B∞ (it is a graph), see Figure 4. This is the classical argument used in [20]. The non-degeneracy condition (2) can be verified on examples. Theorem 5.1. [7] Let ⎧ 2 ⎪ ⎨ a2 u f (x, u) = a3 (x)u3 ⎪ ⎩ a4 u4

a2 = 0  a3  := (1/π ) 0π a3 (x) = 0 a4 = 0 .

(6)

There exist ε0 > 0, σ¯ > 0, a C∞ -curve [0, ε0 )  ε → u(ε ) ∈ Xσ¯ /2,s , a Cantor set C ⊂ [0, ε0 ) of asymptotically full measure, such that, ∀ ε ∈ C , u(ε ) is a solution of equation (1) of section 4 with frequency respectively

408

M. Berti ε ε0

ν1(ε)

Fig. 4 The Cantor set B∞ in which the range equation is solved and the solutions v1 (ε ) of the bifurcation equation (1).

⎧√ 2 ⎪ ⎨ 1 − 2ε ω= 1 − 2ε signa3  ⎪ ⎩√ 1 − 2ε 2 .

(7)

Existence of periodic solutions for completely resonant wave equations like (1) of section 2 has been proved also in [25] if f (u) = u3 + O(u5 ) (to have the nondegeneracy condition) and solving the small divisors problem with the Lindsted series method. Remark 5.1. Under periodic boundary conditions Bourgain [13] proved, when f = u3 + O(u4 ), existence of periodic solutions, bifurcating from exact traveling waves u = δ p0 (ω t + x) of utt − uxx + u3 = 0. More recently Yuan [44] has proved, still for periodic boundary conditions, existence of certain types of quasi-periodic solutions. We remark that the Arnold non-degeneracy condition is generically satisfied in [20] when the bifurcation equation is two dimensional (case of the Lyapunov center theorem), but it is a difficult task yet for partially resonant PDEs like utt − uxx + a1 (x)u = f (x, u) where the bifurcation equation is 2m-dimensional. In this case, considered in [21], the non-degeneracy condition is verified on examples.

6 A variational principle on a Cantor set To relax the Arnold non-degeneracy condition, it is natural to make use of the “Percival reduced variational principle” for solving the bifurcation equation (1) of section 5. The major difficulty, explained at the beginning of Section 5, is to prove the intersection between the solutions of the bifurcation and the range equations for positive measure sets of frequencies. We present below the results and the ideas in [8] where we refer for complete proofs and details.

Variational methods for Hamiltonian PDEs

409

Remark 6.1. New ideas in variational perturbation theory of critical points can shed some light on challenging problems like the generalization of the Weinstein–Moser and Fadell–Rabinowitz theorems for quasi-periodic solutions, see e.g. [19], where the main difficulty arises exactly by a variational principle on a Cantor set. This problem is also related to degenerate KAM theory, see e.g. [38, 39] and references therein. The weak BV-dependence on the frequency. We prove that, if there is a path of solutions ε → v1 (ε ) of (Q1) equation (1) of section 5 which depends just in a BV way, i.e. (1) ∃ d > 0 such that ε0d Var[0,ε0 ] v1 (ε ) ≤ C < +∞ then the Cantor set C defined in (4) of section 5 has asymptotically full measure, i.e. satisfies (5) of section 5. It’s the explicit expression of the Cantor set B∞ in (8) of section 4 to suggest a condition like this (in [8] it is stated in a slightly different way). Note that the variation of ε → v1 (ε ) could be very big. The idea is roughly the following. We have to bound the measure of the complementary set Cc =

/

Sl, j

where

(l, j)∈R

% $  2γ M(ε )   , Sl, j := ε ∈ [0, ε0 ] | ω (ε )l − j + ε < 2j (l + j)τ

/ R := {(l, j) | l = j , l ≥ 1/3ε0 , (1 − 4ε0 )l ≤ j ≤ (1 + 4ε0 )l} (otherwise Sl, j = 0) # ε , v1 (ε ))) satisfies a condition like (1) because M(·, ·) and and M(ε ) := M(v1 (ε ), w( # ·) are smooth. Calling al, j := inf Sl, j , bl, j := sup Sl, j , the measure of each Sl, j w(·, can be bounded like   |M(al, j ) − M(bl, j )| γ . |Sl, j | ≤ C τ +1 + ε0 l jl If all the (al, j , bl, j ) were disjoint, we have to excise all the Sl, j . The measure can be bounded like   /   Sl, j  ≤ C 

∑

(l, j)∈R

(l, j)∈R

γ l τ +1

+ ε0

∑

1 1 ε0 ≤l≤ ε b 0

C +Cε0 jl

∑

l≥ 1b

|M(al, j ) − M(bl, j )| jl

ε0

where also in the second and in the third sum (l, j) ∈ R. The first and the second term are easily shown to satisfy o(ε0 ). The third term can be bounded as Cε0 ε02b

∑

|M(al, j ) − M(bl, j )| ≤ Cε01+2b Var[0,ε0 ] M ≤ ε01+2b ε0−d C = o(ε0 )

l≥1/3ε0b

taking 2b > d. The detailed argument in Section 5.2 of [8] takes into account possible overlapping of the sets (al, j , bl, j ) to show that always | ∪(l, j)∈R Sl, j | ≤ o(ε0 ) + ε01+2b Var[0,ε0 ] M.

410

M. Berti

The weak non-degeneracy condition. We are not able to ensure the BV-property (1) for any f (x, u) = a(x)u p . Therefore we introduce parameter-dependent nonlinearities M

f (λ , x, u) = a(x)u p + ∑ λi bi (x)uqi ,

qi ≥ q¯ > p ≥ 2

(2)

i=1

where q can be arbitrarily large and λi ∈ R are the parameters. We remark that, since qi > p, the nonlinearities λi bi (x)uqi do not change the 0th-order bifurcation equation (3) of section 4, which in particular might have only degenerate solutions. Actually, since the exponents qi can be arbitrarily large, we are adding arbitrarily small corrections bi (x)uqi = o(u p ) for u → 0. The main idea for proving the BV-property (1) for nonlinearities like in (2) is somehow related to the Struwe “monotonicity method” [40] for parameters dependent functionals possessing the mountain pass geometry. We can infer the BV-property for the mountain pass solutions of the bifurcation equation by a BVinformation on the derivatives (w.r.t λ ) of the mountain pass critical levels, choosing properly the exponents qi and the coefficients bi . Normalizing the period and rescaling the amplitude u → δ u, we look for solutions of & ω 2 utt − uxx = ε g(δ , λ , u) (3) u(t, 0) = u(t, π ) = 0 qi −p b (x)uqi . Critical points of where ε := δ p−1 and g(δ , λ , u) := a(x)u p + ∑M i i=1 λi δ the action functional

Ψ (δ , λ , u) :=

 Ω

M ω 2 2 u2x u p+1 ut − + ε a(x) + ∑ λi δ qi −1 2 2 p + 1 i=1

 Ω

bi (x)

uqi +1 qi + 1

are solutions of (3). We perform the same Lyapunov–Schmidt reduction as above. Once the (Q2)-equation and the range equation are solved, the latter on a Cantor set B∞ , we need to find solutions v1 (δ , λ ) of the (Q1)-equation. As above, we need to find a critical point v1 of # (δ , λ , v1 ) = Ψ (δ , λ , v1 + v2 (δ , λ , v1 , w( # δ , λ , v1 )) + w( # δ , λ , v1 )) Φ # can be written Φ # (δ , λ , v1 ) = such that (δ , λ , v1 ) ∈ B∞ . The reduced functional Φ εΦ (δ , λ , v1 ) and Φ (δ , λ , ·) has, ∀δ small, ∀|λ | ≤ 1, a not empty Mountain-Pass critical set K (δ , λ ) ⊂ B(2Rc ;V1 ) \ {0} at the mountain pass critical value c(λ , δ ) = Φ (δ , λ , K (δ , λ )) .

(4)

Furthermore K (δ , λ ) → ΠV1 Kc as δ → 0. The key observation is that the mountain pass value c(δ , λ ) is a semiconcave function, namely c(δ , λ ) − K(δ 2 + |λ |2 ) is concave for some K large enough, see

Variational methods for Hamiltonian PDEs

411

section 2 in [8]. Therefore c(δ , λ ) is differentiable almost everywhere and the derivatives ∂λi c(δ , λ ) are BV functions. Furthermore, at the points where c(λ , δ ) is differentiable,

∂λi c(δ , λ ) = (∂λi Φ )(δ , λ , K (δ , λ )) ,

∀i = 1, . . . , M .

(5)

Formally (5) follows differentiating in (4) since (∂u Φ )(δ , λ , K (δ , λ )) = 0. For each |λ | ≤ 1 we define a path V1 (·, λ ) : [0, δ0 ] → K (δ , λ ) of critical points of Φ (δ , λ , ·). By (5) the functions (∂λi Φ )(δ , λ , V1 (δ , λ )) are BV in the variables (δ , λ ). Hence, for a.e. |λ | ≤ 1, the functions (of one variable)

δ → (∂λi Φ )(δ , λ , V1 (δ , λ )) are BV .

(6)

Q UESTION : How to infer the BV-property (1) for δ → V1 (δ , λ ) from (6)? Differentiating the reduced action functional ! " (∂λi Φ )(δ , λ , v1 ) = δ qi −p Φi (v1 ) + Ri (δ , λ , v1 )

(7)

where Φi : B(2Rc ;V1 ) → R are

Φi (v1 ) :=

1 qi + 1

 Ω


qi +1 bi (x) v1 + v2 (v1 )

with v2 (v1 ) := v2 (0, 0, v1 , 0), and |Ri (δ , λ , v1 )| = O(δ ), |∇v1 Ri (δ , λ , v1 )| = O(δ ). Here the choice of the nonlinearities bi (x)uqi enters into play. The required weak non-degeneracy condition is that the ∇Φi (v1 ) generate V1 , i.e. ∀v1 ∈ ΠV1 Kc ,

span{∇Φi (v1 ) , i = 1, . . . , M} ≡ V1 .

(8)

In this case, in any neighborhood of ΠV1 Kc a finite set of Φi (v1 ) can be seen as a local chart of coordinates. Hence, by the BV dependence (6), and (7) we infer the (BV) property for the path of critical points δ → V1 (δ , λ ). In [8], the weak nondegeneracy condition (8) is verified, proving the following theorem: Theorem 6.1. [8] For any q¯ > p there exist M ∈ N, integer exponents q¯ ≤ q1 ≤ . . . ≤ qM and coefficients b1 , . . . ,bM ∈ H 1 (0, π ) depending only on a(x), such that for almost every parameter λ = (λ1 , . . . , λM ), |λ | ≤ 1, equation (1) of section 2 with nonlinearity f (λ , x, u) like in (2) possesses periodic solutions for an asymptotically full measure Cantor set of frequencies ω close to 1. Furthermore, given a(x)u p , bi (x)uqi , Theorem 6.1 is valid also adding any nonlinear term r(x, u) = ∑k>p rk (x)uk , with ∑k>p rk H 1 ρ k < +∞ for some ρ > 0. The

412

M. Berti

term r has an influence only on the full measure set of parameters λ for which the existence result holds (in the previous argument we use just the derivatives w.r.t. λ ). Theorem 6.1 can be interpreted as a genericity result in the sense of Lebesgue measure, see for details [8].

7 Forced vibrations In this section we look for T -periodic solutions of & utt − uxx = ε f (t, x, u) u(t, 0) = u(t, π ) = 0

(1)

when the nonlinearity is T -periodic, namely f (t + T, x, u) = f (t, x, u), ∀t. We suppose that the forcing frequency ω := 2π /T is a rational number, for simplicity

ω = 1,

i.e. T = 2π

(if ω ∈ / Q a small divisor problem similar to the one discussed in the previous sections appears, see e.g. [2] and references therein). The spectrum of the D’Alembert operator in a space of 2π -periodic functions is (see (3) of section 2)

σ (∂tt − ∂xx ) = − l 2 + j2 | l ∈ Z , j ∈ N ⊂ Z . Therefore zero is an eigenvalue of infinite multiplicity (when |l| = j) but the other eigenvalues are well separated from zero if |l| = j. For this reason the difficulty is not solving the range equation, but the bifurcation equation which has an intrinsic lack of compactness. The first breakthrough regarding this problem was achieved by Rabinowitz [34] under the strong monotonicity assumption (∂u f )(t, x, u) ≥ β > 0 and in [35] for weakly monotone nonlinearities like f (t, x, u) = u2k+1 + G(t, x, u) where G(t, x, u2 ) ≥ G(t, x, u1 ) if u2 ≥ u1 . For several other results in the monotone case see e.g. [16] and references therein. The monotonicity of f is deeply exploited to compensate the for lack of compactness in the infinite dimensional bifurcation equation. Little is known without the monotonicity. Willem [43], Hofer [26] and Coron [17] have proved some existence result for nonlinearities like f (t, x, u) = g(u) + h(t, x), where g(u) satisfies suitable linear growth conditions, ε = 1, and under additional symmetries or non-resonance assumptions. We now present the recent existence results of [5] for non-monotone nonlinearities. This will highlight a completely different use of variational methods.

Variational methods for Hamiltonian PDEs

413

7.1 The variational Lyapunov–Schmidt reduction In view of the variational argument that we shall use to solve the bifurcation equa1/2 tion we look for solutions u of (1) in the Banach space E := H 1 (Ω )∩C0 (Ω¯ ) where 1/2 H 1 (Ω ) is the usual Sobolev space, Ω := T × [0, π ], and C0 (Ω¯ ) is the space of all the 1/2-Hölder continuous functions satisfying u(t, 0) = u(t, π ) = 0 endowed with norm uE := uH 1 (Ω ) + uC1/2 (Ω¯ ) . Critical points of the Lagrangian action functional Ψ ∈ C1 (E, R)

Ψ (u) :=

  2 u

 u2x − + ε F(t, x, u) dtdx 2 2 t

Ω

where (∂u F)(t, x, u) = f (t, x, u) are weak solutions of (1). For ε = 0, the critical points of Ψ in E reduce to the space V := N ∩ H 1 (Ω ) ⊂ E where

(2)

% $  2 vˆ = 0 N := v = v(t ˆ + x) − v(t ˆ − x) =: v+ − v− : vˆ ∈ L (T) and T

is the L2 -closure of the classical solutions of equation (2) of section 2. We have V ⊂ E because any vˆ ∈ H 1 (T) is 1/2-Hölder continuous. The only difference with respect to the space V introduced in (2) of section 3 is that the functions v in (2) are not necessarily even in time. Theorem 7.1. [5] Let f (t, x, u) = u2k + h(t, x) where h ∈ N ⊥ satisfies h(t, x) > 0 a.e. in Ω . Then ∀ε small enough, there exists at least one weak solution u ∈ E of (1) with uE ≤ C|ε |. Theorem 7.1 is a particular case of a more general result which holds without any growth condition for f , see Theorems 1, 2 in [5]. Moreover, the solution u is proved to be more regular, when h is more regular. This is somewhat surprising: Brezis–Nirenberg [16] and Rabinowitz [36] have proved regularity of solutions if f is strictly monotone in u. For example, yet the solutions in [35] are only continuous functions. The less regular part of the solution is the component in V because of the lack of compactness. / N ⊥ , periodic Remark 7.1. The assumption h ∈ N ⊥ is not of technical nature: if h ∈ solutions of (1) do not exist in any fixed ball {uL∞ ≤ R} for ε small. To prove Theorem 7.1 we perform a Lyapunov–Schmidt reduction, decomposing E = V ⊕W and N ⊥ := {h ∈ L2 (Ω ) |

 Ω

where

hv = 0, ∀v ∈ N}.

W := N ⊥ ∩ E

414

M. Berti

Projecting (1), for u = v + w with v ∈ V , w ∈ W , yields & 0 = ΠN f (v + w) bifurcation equation range equation wtt − wxx = εΠN ⊥ f (v + w)

(3)

where ΠN and ΠN ⊥ are the projectors from L2 (Ω ) onto N and N ⊥ . The range equation. The inverse of the D’Alembert operator −1 : N ⊥ → W defined by fl j −1 f := ∑ eilt sin jx , ∀ f ∈ N⊥ 2 + j2 −l j≥1, j =|l| is a bounded operator, i.e. −1 f E ≤ C  f L2 , see [16]. Fixed points w ∈ W of w = ε −1 ΠN ⊥ f (v + w) are solutions of the range equation. By the contraction mapping theorem we have: Lemma 7.1. (Solution of the range equation) ∀ R > 0, ∃ ε0 (R) > 0, C0 (R) > 0, such that ∀ |ε | ≤ ε0 (R) and ∀ v ∈ N with vL∞ ≤ 2R there exists a unique solution w(ε , v) ∈ W of the range equation satisfying w(ε , v)E ≤ C0 (R)|ε | .

(4)

Moreover the map (ε , v) → w(ε , v) is C1 ({vL∞ ≤ 2R},W ).

7.2 The bifurcation equation There remains the infinite dimensional bifurcation equation

ΠN f (v + w(ε , v)) = 0 which is the Euler-Lagrange equation of the reduced action functional

Φε : {vH 1 < 2R} → R ,

Φε (v) := Ψ (v + w(ε , v)) .

Φε can be written (like in (6) of section 3 with w = 1) as  ! " 1 F(v + w(ε , v)) − f (v + w(ε , v))w(ε , v) dt dx . Φε (v) = ε 2 Ω To find critical points of Φε , we make a constrained minimization like in [34] (we don’t have compactness to try any critical point theory). By the compact embedding (V,  · H 1 ) → (V,  · ∞ ) (see Remark 3.1) it easy to deduce that Lemma 7.2. ∀R > 0, Φε attains a minimum v¯ in B¯ R := {v ∈ V, vH 1 ≤ R} .

Variational methods for Hamiltonian PDEs

415

Fig. 5 The variational inequality.

If v¯ ∈ ∂ B¯ R then we deduce only the variational inequality ¯ φ] = Dv Φε (v)[

 Ω

f (v¯ + w(ε , v)) ¯ φ ≤0

(5)

¯ φ H 1 > 0, see Figure 5. for any admissible variation φ ∈ V such that v, The heart of the existence proof is to obtain, choosing suitable admissible varia¯ H 1 < R∗ for some R∗ > 0 independent of ε . tions, the a-priori estimate v It is here where the monotonicity of f plays its role to prove Rabinowitz’s theorem [34]. On the other hand, the difficulty for dealing with non-monotone nonlinearities is well highlighted by the nonlinearities f = u2k + h(t, x). In this case the variational inequality (5) vanishes identically for ε = 0, because  Ω

(v¯2k + h(t, x))φ ≡ 0,

since h ∈ N ⊥ and the following lemma. Lemma 7.3. [5] If v1 , . . . , v2k+1 ∈ V then then v2k ∈ W .



∀φ ∈ V

Ω v1 · . . .· v2k+1

= 0. In particular, if v ∈ V

Therefore, for deriving, if possible, the required a priori estimates, we have to develop the variational inequality (5) at higher orders in ε . It is convenient to perform in (1) with f = u2k + h the change of variables u → ε (H + u) where H is a weak solution of ⎧ ⎪ ⎨ Htt − Hxx = h (6) H(t, 0) = H(t, π ) = 0 ⎪ ⎩ H(t + 2π , x) = H(t, x) and, since h > 0 in Ω , we can always choose H(t, x) > 0, ∀(t, x) ∈ Ω (see the “maximum principle” theorem proved in [5]). Therefore (renaming ε 2k → ε ) we look for 2π periodic solutions of & utt − uxx = ε f (t, x, u) where f (t, x, u) := (H + u)2k . (7) u(t, 0) = u(t, π ) = 0 Implementing a variational Lyapunov–Schmidt reduction as above, we find the existence of a constrained minimum v¯ ∈ B¯ R and the variational inequality

416

M. Berti

 Ω

(H + v¯ + w(ε , v)) ¯ 2k φ ≤ 0 .

(8)

The required a priori estimate for the H 1 -norm of v¯ is proved in several steps inserting into the variational inequality (8) suitable admissible variations. We consider only the more difficult case k ≥ 2. The following key “coercivity” estimate is heavily exploited. Lemma 7.4. (Coercivity estimate) [5] Let H ∈ C(Ω¯ ) with H > 0 in Ω . Then  Ω

Hv2k ≥ ck (H)

 Ω

where ck (H) := 4−k minΩ¯ α H > 0, αk := k

∀ v ∈ N ∩ L2k (Ω )

v2k ,

1 4(1+2k) ,

(9)

and Ωα := T × (απ , π − απ ).

The inequality (9) is not trivial because H vanishes at the boundary ∂ Ω (i.e. ˆ + x)− v(t ˆ − x) is the superpoH(t, 0) = H(t, π ) = 0). It holds true because v = v(t sition of waves which “spend a lot of time” far from x = 0 and x = π . In the sequel κi will denote positive constants independent of ε . Step 1: the L2k -estimate. Insert φ := v¯ in (8). φ is an admissible variation since ¯ 2H 1 > 0. Setting w¯ := w(ε , v) ¯ v, ¯ φ H 1 = v  Ω



(v¯ + H)2k v¯ = (8)

≤

Ω



Ω



(v¯ + H + w) ¯ 2k v¯ +

Ω

(4)

[(v¯ + H)2k − (v¯ + w¯ + H)2k ]v¯ ≤ |ε |C1 (R)v ¯ L1 ≤ 1 (10)

for |ε | ≤ ε1 (R) where 0 < ε1 (·) ≤ ε0 (·). Since (10) 

1 ≥

Ω

[(v¯ + H)2k − (v¯ + w¯ + H)2k ]v¯

2k

 Ω

v¯2k+1 = 0 by Lemma 7.3,

 !

" (v¯ + H)2k − v¯2k v¯ Ω   2k−2  2k 2k v¯ j+1 H 2k− j 2kH v¯ + ∑ = j Ω j=0

(v¯ + H) v¯ =

(11)

(9)

≥ 2kck (H)v ¯ 2k − κ1 v ¯ 2k−1 − κ2 v ¯ L2k L2k L2k

using Hölder inequality to estimate v ¯ Li ≤ Ci,k v ¯ L2k (i ≤ 2k − 1). By (11) v ¯ L2k ≤ κ3

for

|ε | ≤ ε1 (R) .

(12)

ˆ¯ + x)) − q(v(t ˆ¯ − x)) ∈ V where Step 2: the L∞ -estimate. We now choose φ = q(v(t ⎧ if |λ | ≤ M ⎨0 q(λ ) := λ − M if λ ≥ M ⎩ λ + M if λ ≤ M

Variational methods for Hamiltonian PDEs

417

and

1 ˆ¯ L∞ (T) . M := v 2 We can assume M > 0, i.e. v¯ is not identically zero.

(13)

Lemma 7.5. ([34]) φ ∈ V is an admissible variation and v(t, ¯ x)φ (t, x) ≥ 0, ∀(t, x) ∈ Ω . Using (8), setting w¯ := w(ε , v), ¯  Ω

(v¯ + H)2k φ = (8)

≤

 Ω



Ω

(v¯ + H + w) ¯ 2k φ +

 Ω

[(v¯ + H)2k − (v¯ + w¯ + H)2k ]φ

[(v¯ + H)2k − (v¯ + w¯ + H)2k ]φ

(4)

≤ |ε |C2 (R)φ L1 ≤ φ L1

(14)

for |ε | ≤ ε2 (R) where 0 < ε2 (·) ≤ ε1 (·). Now, using that (14) 

φ L1 ≥

(v¯ + H) φ = 2k

Ω

≥ 

which implies, since  Ω1/4

Ω

 ! Ω

 Ω

v¯2k φ = 0 by Lemma 7.3,

" (v¯ + H)2k − v¯2k φ

2kH v¯2k−1 φ − κ4 (v ¯ 2k−2 L∞ + 1)φ L1

Ω



H v¯2k−1 φ ≥ κ5

2k−1 Ω1/4 φ v¯

(because v¯φ ≥ 0), that

ˆ¯ L1 (T) v¯2k−1 φ ≤ κ6 (v ¯ 2k−2 ¯ 2k−2 L∞ (Ω ) + 1)φ L1 (Ω ) ≤ κ7 (v L∞ (Ω ) + 1)q(v)

ˆ¯ L1 (T) . We have to give a lower bound of since φ L1 (Ω ) ≤ q(v)  2k−1

Ω1/4

v¯

φ=



(v¯φ )v¯

2k−2

Ω1/4



=

Ω1/4

(v¯φ )(v¯+ − v¯− )2k−2 .

By the elementary inequality (a−b)2k ≥ a2k +b2k −2k(a2k−1 b+ab2k−1 ), ∀a, b ∈ R,   ! " v¯2k−1 φ ≥ v¯φ v¯2k−2 + v¯2k−2 − (2k − 2)(v¯2k−3 v¯− + v¯+ v¯2k−3 ) + − + − Ω1/4

Ω1/4



=2

Ω1/4

v¯2k−1 q+ − v¯2k−1 q− + v¯2k−2 v¯− q− − v¯2k−2 v¯− q+ + + + +

+ (2k − 2)[−v¯2k−2 v¯− q+ + v¯2k−2 v¯− q− − v¯2k−3 v¯2− q− + v¯2k−3 v¯2− q+ ] + + + + ≥2 −2



Ω1/4



Ω1/4

v¯2k−1 q+ +

(15)

v¯2k−1 q− + (2k − 1)v¯2k−2 v¯− q+ + (2k − 2)v¯2k−3 v¯2− q− (16) + + +

418

M. Berti

where in the last inequality we used that v¯+ q+ , v¯− q− ≥ 0 (since λ q(λ ) ≥ 0) and so v¯2−q+ ≥ 0. The dominant term (15) is estimated using the identity v¯2k−2 v¯− q− , v¯2k−3 + +  1 p(t + x) dt dx = p(t − x) dt dx = π (1 − 2 α ) Ωα Ωα T p(s)ds, ∀p ∈ L (T), 

2

2 v¯2k−1 q+ = 2π (1 − ) + 4 Ω1/4

 T

ˆ¯ ˆ¯ L1 (T) ≥ π M 2k−1 q(v) v¯ˆ2k−1 (s)q(v(s))ds

(17)

because λ 2k−1 q(λ ) ≥ M 2k−1 |q(λ )|. The terms in (16) are estimated by             ˆ¯ 2k−1 ˆ¯ L1 (T) v¯2k−1 q−  ≤ 2 v¯2k−1 q(v) 2  |q− | ≤ v + + L2k−1 (T)  Ω1/4  Ω    
    2k−2 ˆ¯ L1 (T) v¯ q+ v¯−  ≤ v¯ˆ2k−2 q(v) ¯ˆ L1 (T) v ¯ˆ L1 (T) ≤ (2M)2k−2 q(v) ¯ˆ L1 (T) v 2  Ω1/4 +     
   ˆ¯ L1 (T) ≤ (2M)2 v ˆ¯ 2k−3 ˆ¯ L1 (T) v¯2k−3 v¯2− q−  ≤ vˆ¯2k−3 L1 (T) vˆ¯2 q(v) q(v) 2 L2k−3 (T)   Ω1/4 + By the previous inequalities, (17), Hölder inequality and (12)  Ω1/4

ˆ¯ L1 (T) − κ7 (M 2k−2 + 1)q(v) ˆ¯ L1 (T) . v¯2k−1 φ ≥ π M 2k−1 q(v)

(18)

By (15) and (18)
 2k−2 ˆ¯ L1 (T) ≤ κ8 v ˆ¯ L1 (T) M 2k−1 q(v) ¯ 2k−2 + M + 1 q(v) ∞ L (Ω )
 ˆ¯ L1 (T) ≤ κ9 M 2k−2 + 1 q(v) ˆ¯ L∞ (T) (see (13)) and finally M 2k−1 ≤ κ9 (M 2k−2 + 1). Since M := 12 v v ¯ L∞ ≤ κ10

for

|ε | ≤ ε2 (R) .

Step 3: The H1 -estimate. The H 1 -estimate is carried out inserting in (8) the variation φ := −D−h Dh v¯ where (Dh f )(t, x) :=

f (t + h, x) − f (t, x) h

is the difference quotient with respect to t. Note that φ is admissible because ¯ v ¯ H 1 = Dh v, ¯ Dh v ¯ H 1 > 0. With arguments similar to the previous ones, −D−h Dh v, we can deduce that, for some 0 < ε3 (R) ≤ ε2 (R), v ¯ H 1 < κ11 ,

∀ |ε | ≤ ε3 (R) .

P ROOF OF T HEOREM 7.1 COMPLETED . Take R∗ := κ11 and ε∗ := ε3 (R∗ ). There¯ ε )H 1 < R∗ is an interior minimum of Φε in BR∗ . fore, ∀ |ε | ≤ ε∗ , v(

Variational methods for Hamiltonian PDEs

419

Acknowledgements I wish to thank P. Baldi, L. Biasco, P. Bolle and M. Procesi for sharing much work together, as well as D. Bambusi, W. Craig and P. Rabinowitz for many useful discussions.

References 1. Ambrosetti A., Rabinowitz P., Dual Variational Methods in Critical Point Theory and Applications, Journ. Func. Anal, 14, 349–381, 1973. 2. Baldi P., Berti M., Forced vibrations of a nonhomogeneous string, to appear in SIAM Journal on Mathematical Analysis. preprint 2006. 3. Bambusi D., Paleari S., Families of periodic solutions of resonant PDEs, J. Non. Sci., 11, 69–87, 2001. 4. Berti M., Topics on Nonlinear oscillations of Hamiltonian PDEs, Progress in Nonlinear Differential Equations and Its Applications, Vol. 74, Birkhauser, Boston, 2008, Brezis. Ed. 5. Berti M., Biasco L., Forced vibrations of wave equations with non-monotone nonlinearities, Annales de l’Institute H. Poincare’, Analyse nonlinaire, 23, 4, 437–474, 2006. 6. Berti M., Bolle P., Periodic solutions of nonlinear wave equations with general nonlinearities, Comm. Math. Phys. Vol. 243, 2, 315–328, 2003. 7. Berti M., Bolle P., Cantor families of periodic solutions for completely resonant nonlinear wave equations, Duke Mathematical Journal, 134, 2, 359–419, 2006. 8. Berti M., Bolle P., Cantor families of periodic solutions for wave equations via a variational principle, to appear in Advances in Mathematics. preprint 2006. 9. Berti M., Bolle P., Cantor families of periodic solutions of wave equations with Ck nonlinearities, to appear in Nonlinear Differential Equations and Applications. preprint 2007. 10. Berti M., Procesi M., Quasi-periodic solutions of completely resonant forced wave equations, Communications in Partial Differential Equations, 31, 959–985, 2006. 11. Bourgain J., Construction of quasi-periodic solutions for Hamiltonian perturbations of linear equations and applications to nonlinear PDE, Internat. Math. Res. Notices, no. 11, 1994. 12. Bourgain J., Quasi-periodic solutions of Hamiltonian perturbations of 2D linear Schrödinger equations, Ann. of Math., 148, 363–439, 1998. 13. Bourgain J., Periodic solutions of nonlinear wave equations, Harmonic analysis and partial differential equations (Chicago, IL, 1996), 69–97, Chicago Lectures in Math., Univ. Chicago Press, Chicago, IL, 1999. 14. Bourgain J., Green’s function estimates for lattice Schrödinger operators and applications, Annals of Mathematics Studies, 158. Princeton University Press, Princeton, NJ, 2005. 15. Brezis, H., Coron, J.-M., Nirenberg, L., Free vibrations for a nonlinear wave equation and a Theorem of P. Rabinowitz, Comm. Pure Appl. Math. 33, no. 5, 667–684, 1980. 16. Brezis H., Nirenberg L., Forced vibrations for a nonlinear wave equation, Comm. Pure Appl. Math. 31, no. 1, 1–30, 1978. 17. Coron J.-M., Periodic solutions of a nonlinear wave equation without assumption of monotonicity, Math. Ann. 262, no. 2, 273–285, 1983. 18. Craig W., Problèmes de petits diviseurs dans les équations aux dérivées partielles, Panoramas et Synthèses, 9, Société Mathématique de France, Paris, 2000. 19. Craig W., Invariant tori for Hamiltonian PDE, Nonlinear dynamics and evolution equations, 53-66, Fields Inst., 48, 2006, AMS-Providence. 20. Craig W., Wayne E., Newton’s method and periodic solutions of nonlinear wave equation, Comm. Pure and Appl. Math, vol. XLVI, 1409–1498, 1993. 21. Craig W., Wayne E., Nonlinear waves and the 1 : 1 : 2 resonance, Singular limits of dispersive waves, 297–313, NATO Adv. Sci. Inst. Ser. B Phys., 320, Plenum, New York, 1994. 22. De La Llave R., Variational Methods for quasi-periodic solutions of partial differential equations, World Sci. Monogr. Ser. Math. 6, 214–228, 2000. 23. Eliasson H., Kuksin S., KAM for nonlinear Schrödinger equation, to appear in Annals of Mathematics. preprint 2006.

420

M. Berti

24. Fadell E., Rabinowitz P., Generalized cohomological index theories for the group actions with an application to bifurcation questions for Hamiltonian systems, Inv. Math. 45, 139–174, 1978. 25. Gentile G., Mastropietro V., Procesi M., Periodic solutions for completely resonant nonlinear wave equations, Comm. Math. Phys. 256, no. 2, 437–490, 2005. 26. Hofer H., On the range of a wave operator with non-monotone nonlinearity, Math. Nachr. 106, 327–340, 1982. 27. Lidskij B. V., Shulman E. I., Periodic solutions of the equation utt − uxx + u3 = 0, Funct. Anal. Appl. 22, 332–333, 1988. 28. Kuksin S., Hamiltonian perturbations of infinite-dimensional linear systems with imaginary spectrum, Funktsional. Anal. i Prilozhen. 21, no. 3, 22–37, 95, 1987. 29. Kuksin S., Analysis of Hamiltonian PDEs, Oxford Lecture Series in Mathematics and its Applications, 19. Oxford University Press, New York, NY 2000. 30. Kuksin S., Pöschel J., Invariant Cantor manifolds of quasi-periodic oscillations for a nonlinear Schrödinger equation, Ann. of Math, 2, 143, no. 1, 149–179, 1996. 31. Moser J., Periodic orbits near an Equilibrium and a Theorem by Alan Weinstein, Comm. Pure Appl. Math., vol. XXIX, 1976. 32. Moser J., Zehnder E., Notes on Dynamical Systems, Courant Lecture Notes 2005. 33. Pöschel J., Quasi-periodic solutions for a nonlinear wave equation, Comment. Math. Helv., 71, no. 2, 269–296, 1996. 34. Rabinowitz P., Periodic solutions of nonlinear hyperbolic partial differential equations, Comm. Pure Appl. Math., 20, 145–205, 1967. 35. Rabinowitz P., Time periodic solutions of nonlinear wave equations, Manusc. Math. 5, 165– 194, 1971. 36. Rabinowitz P., Free vibrations for a semilinear wave equation, Comm. Pure Appl. Math. 31, no. 1, 31–68, 1978. 37. Rabinowitz P., Minimax methods in critical point theory with applications to differential equations, CBMS Regional Conference Series in Mathematics, 65. Published for the Conference Board of the Mathematical Sciences, Washington, DC; by the American Mathematical Society, Providence, RI, 1986. 38. Sevryuk M.B., The lack-of-parameters problem in the KAM theory revisited, in Hamiltonian systems with three or more degrees of freedom, ed. C. Simó, NATO ASI Series C, Math Phys. Sci, vol. 533, Kluwer, Dordrecht, The Netherlands, 1999, 568–572. 39. Sevryuk M.B., The classical KAM theory at the dawn of the twenty-first century, Mosc. Math. J. 4 (3), 1113–1144. 40. M. Struwe, The existence of surfaces of constant mean curvature with free boundaries, Acta Math., 1988, 19–64. 41. Wayne E., Periodic and quasi-periodic solutions of nonlinear wave equations via KAM theory, Commun. Math. Phys. 127, no. 3, 479–528, 1990. 42. Weinstein A., Normal modes for Nonlinear Hamiltonian Systems, Inv. Math, 20, 47–57, 1973. 43. Willem M., Density of the range of potential operators, Proc. Amer. Math. Soc. 83, 2, 341– 344, 1981. 44. Yuan X., Quasi-periodic solutions of completely resonant nonlinear wave equations, J. Diff. Equations, 230, no. 1, 213–274, 2006. 45. Yuan X., A KAM theorem with applications to partial differential equations of higher dimensions, Comm. Math. Phys. 275 (2007), no. 1, 97–137.

Spectral gaps of potentials in weighted Sobolev spaces Jürgen Pöschel1

Abstract We consider the Schr¨odinger operator L = −d 2 / dx2 + q on the interval [0, 1] depending on an L2 -potential q and endowed with periodic or anti-periodic boundary conditions. We prove results about correspondencies between the asymptotic behaviour of the spectral gaps of L and the regularity of q in the Gevrey case, among others. The proofs are based on a Fourier block decomposition due to Kappeler & Mityagin, and a novel application of the implicit function theorem.

1 Results We consider the Schr¨odinger operator L=−

d2 +q dx2

on the interval [0, 1] depending on an L2 -potential q and endowed with periodic or anti-periodic boundary conditions. In this case, L is also known as Hill’s operator. Its spectrum is pure point, and for real q consists of an unbounded sequence of real periodic eigenvalues

λ0+ (q) < λ1− (q)
λ1+ (q) < · · · < λn− (q)
λn+ (q) < · · · . Their asymptotic behaviour is

λn± = n2 π 2 + [q] + 2 (n),

1 Institut

für Analysis, Dynamik und Optimierung, Universität Stuttgart, Pfaffenwaldring 57 D-70569 Stuttgart e-mail: [email protected]

W. Craig (ed.), Hamiltonian Dynamical Systems and Applications, 421–430. c 2008 Springer Science + Business Media B.V.


421

422

J. Pöschel

where [q] denotes the mean value of q and 2 (n) a generic square sumable term. Equality may occur in every place with a ‘
’-sign, and one speaks of the gap lengths

γn (q) = λn+ (q) − λn− (q),

n  1,

of the potential q. For complex q, the periodic eigenvalues are still well defined, but in general not real, since L is no longer self-adjoint. Their asymptotic behaviour is the same, however, and we may order them lexicographically – first by their real part, then by their imaginary part – so that

λ0 (q) ≺ λ1− (q) λ1+ (q) ≺ · · · ≺ λn− (q) λn+ (q) ≺ · · · . The gap lengths are then defined as before, but are now complex valued in general.

Classical Results We are interested in the relationship between the regularity of a potential and the sequence of its gap lengths. Hochstadt [5] observed that q ∈ C∞ (S1 , R) ⇔ γn (q) = O(n−k ) for all k  0, and Marˇcenko & Ostrowsk˘ı [27] subsequently showed that q ∈ H m (S1 , R) ⇔

∑ n2m γn2 (q) < ∞

n1

for all nonnegative integers m. Trubowitz [32] then proved that q ∈ Cω (S1 , R) ⇔ γn (q) = O(ε −an ) for some a > 0. Later, due to the realization of the periodic KdV flow as an isospectral deformation of Hill’s operator, other regularity classes such as Gevrey functions and non-real potentials came into focus. Recent results in this direction appear example in [2, 11, 12, 16, 17]. Within certain limits, one may think of the gap lengths as another kind of Fourier coefficients of the potential. It is the purpose of this note to describe some of these recent developments. For more detailed statements and proofs we refer to [10] and later [3].

Weighted Sobolev Spaces As in [16, 17] a weight is a normalized, symmetric and submultiplicative function w : Z → R. That is, for all integers n and m, we have wn  1,

w−n = wn ,

wn+m
wn wm .

Spectral gaps of potentials in weighted Sobolev spaces

423

Within the standard Sobolev space H0 = L2 (S1 , C) of square-integrable functions u = ∑n∈Z un ε 2π inx we then define the weighted Sobolev spaces Hw = {u ∈ H0 : u2w := ∑n∈Z w2n |un |2 < ∞}. To give some examples, let n = 1 + |n|. The Sobolev weights nr , r  0, give rise to the usual Sobolev spaces Hr of 1-periodic, complex-valued functions. The Abel weights1 nr ε a|n| with a > 0 define spaces Hr,a of functions in Hr , which are analytic on the complex strip |Im z| < a/2π and have traces in Hr on the boundary lines. The Gevrey weights σ

wn = nr ε a|n| ,

r  0, a > 0, 0 < σ < 1,

lie in between and give rise to the so called Gevrey spaces Hr,a,σ of smooth functions. Obviously, Hr,a = Hr,a,1  Hr,a,σ  Hr,a,0 = Hr . Yet another weight is for example wn = nr exp




 a|n| , α 1 + log n

α > 0.

Since log wn is subadditive and nonnegative, the limit log wn n→∞ n

χ (w) := lim

exists and is nonnegative [29, no. 98]. Naturally, we call a weight w exponential, if χ (w) > 0. We call w subexponential, if χ (w) = 0 with n−1 log wn converging to zero in an eventually monotone manner. This is not an exact dichotomy, but we are not aware of any interesting weight that does not belong to either class. Clearly, Abel weights are exponential, while Sobolev and Gevrey weights are subexponential.

The Theorems Let hw = {u = (un )n1 ∈ 2 : ∑n1 w2n |un |2 < ∞}, and γ (q) = (γn (q))n1 .

1

The term Abel weights is chosen to go along with Sobolev and Gevrey weights and has no deeper meaning.

424

J. Pöschel

Theorem 1 If q ∈ Hw , then γ (q) ∈ hw . In particular,

∑ w2n |γn (q)|2
9TN q2w +

nN

288 q4w N

for all N  4qw , where TN q = ∑|n|N qn ε 2nπ ix . There is no one-to-one converse to Theorem 1 for exponential weights, as there exist real analytic finite gap potentials such as the Weierstrass ℘-function, which are not entire functions. In this case, fixing any r > 0, we have (γn (u)) ∈ hr,a for all a > 0, but only u ∈ Hr,a for a < a0 . There is, however, a true converse for subexponential weights. We first consider the real case. Theorem 2 Suppose q ∈ H0 is real and γ (q) ∈ hw . If w is subexponential, then q ∈ Hw . If w is exponential, then q is real analytic. Corollary 3 If q is real and w is subexponential, then q ∈ Hw ⇔ γ (q) ∈ hw . For complex potentials, the spectral gap lengths alone do not suffice to determine the regularity of a potential. For instance, Gasymov [13] showed that all gap lengths vanish for complex potentials of the form  q = ∑ qn ε 2nπ ix = ∑ qn zn z=ε 2π ix . n1

n1

But Sansuc & Tkachenko [11] noted that the situation can be remedied by taking into account additional spectral data. In particular, they considered the quantities

δn = µn − τn = µn − (λn+ − λn− )/2, where µn denotes the n-th Dirichlet eigenvalue. The quantities

Γn = |γn | + |δn |, may be considered as a measure of the size of the spectral triangle formed by the points λn− , δn and λn+ . Note that γn
Γn
2γn for real potentials. We then have the following converse theorem. Theorem 4 Suppose q ∈ H0 is real or complex and Γ (q) ∈ hw . If w is subexponential, then q ∈ Hw . If w is exponential, then q is real analytic. Corollary 5 If w is subexponential, then q ∈ Hw ⇔ Γ (q) ∈ hw . For the sake of brevity and simplicity, we will describe the line of reasoning for the real case. For more details, complete proofs and the complex case we refer to [10] and also [3].

Spectral gaps of potentials in weighted Sobolev spaces

425

2 Reduction The idea of the proof of Theorem 1 is due to Kappeler & Mityagin [17]. They employ a Lyapunov–Schmidt reduction scheme called Fourier block decomposition. The aim is to determine those λ near n2 π 2 with n sufficiently large, for which the equation −y

+ qy = λ y admits a nontrivial 2-periodic solution f . As q can be considered small for large n, one expects its dominant modes to be ε ±nπ ix . So it makes sense to separate these modes from the other ones by a Lyapunov–Schmidt reduction. To this end we consider a similarly defined space Hw of 2-periodic functions, and write Hw = Pn ⊕ Qn = span {en , e−n } ⊕ span {ek : |k| = n}, where ek = ε kπ ix . The projections onto Pn and Qn are denoted by Pn and Qn , respectively. With f = u + v = Pn f + Qn f , Hill’s equation decomposes into the so called P- and Q-equations Aλ u = PnV (u + v), Aλ v = QnV (u + v), where Aλ f = f

+ λ f and V f = q f . The operator Aλ has a compact inverse on Qn , when λ is near n2 π 2 . Indeed, this holds on the complex strips Un = { λ : |Re λ − n2 π 2 |
12n} for n  1. Lemma 1 For q ∈ Hw and λ ∈ Un , the operator Tn = VA−1 λ Qn exists and is bounded on Hw with norm 2 Tn w
qw . n Left-multiplying the Q-equation with VA−1 λ we obtain V v = TnVu + TnV v. For n large enough, Tn is a contraction on Hw , and there is a unique solution V v = Tˆn TnVu,

Tˆn = (I − Tn )−1 .

Inserted into the P-equation we get Aλ u = PnVu + Pn Tˆn TnVu = Pn TˆnVu. So the Pand Q-equation reduce to the S-equation Sn u = 0,

Sn = Aλ − Pn TˆnV.

Since Pn is two-dimensional with basis en , e−n , we have the matrix representation   λ − n2 π 2 − an −cn Sn = , −c−n λ − n2 π 2 − an

426

J. Pöschel

with an = TˆnVen ,en ,

cn = TˆnVe−n ,en .

Any nontrivial solution u gives rise to a 2-periodic solution of Aλ f = V f , and vice versa. Hence the following holds. Lemma 2 A complex number λ near n2 π 2 is a periodic eigenvalue of q if and only if the determinant of Sn vanishes. Moreover, from the representations of an and cn one easily obtains the following facts, which we need later. Lemma 3 For n  4qw and λ ∈ Un , 4 |an − q0 |Un , wn |cn − qn |Un , wn |c−n − q−n |Un
q2w . n Moreover, these coefficients are analytic functions of λ and q.

3 Gap Estimates With the preceding results the forward problem of estimating the gap lengths of a potential is fairly straightforward. The determinant of Sn is the quadratic polynomial det Sn = (λ − n2 π 2 − an )2 − |cn |2 , and the distance of its two roots has to be of the order of |cn |. Lemma 4 For n  4qw the determinant of Sn has exactly two roots ξn± in Un , which are contained in the disc |λ − n2 π 2 |
6qw and satisfy  +  ξn − ξn− 2
9|cn c−n | . Un A counting argument then shows that these two roots have to be the two eigenvalues λn± . Consequently, we obtain 2  144 |γn |2 = ξn+ − ξn− 
9|cn c−n |Un
9|qn |2 + 9|q−n |2 + 2 2 q4w n wn by Lemma 3. Multiplying by w2n and summing over n  N we obtain Theorem 1.

4 Coefficient Estimates We now turn to the more subtle problem of estimating the asymptotic behaviour of the Fourier coefficients of a potential in terms of its gap lengths. The geometric

Spectral gaps of potentials in weighted Sobolev spaces

427

aspect is rather straightforward, at least in the real case. The off diagonal elements of Sn have to be bounded in terms of the gap lengths, that is |cn |
|γn |,

n , 1,

where the dot stands for some implicit constant. The identity cn = TˆnVe−n ,en  then leads to an infinite dimensional system of nonlinear equations n = 0,

cn = qn + O2 (. . . , qk , . . . ),

which allows us to bound the qn in terms of the cn and hence in terms of the γn . See [12] for the lengthy details of this argument. Here we present a functional analytic approach based on the fact that the coefficients of Sn – which contain all the necessary data – are analytic functions of q. Indeed, it even suffices to consider Sn at that value of λ where its diagonal vanishes. For m  1 and any weight w we introduce the centered balls Bwm = {q ∈ Hw : qw
m/4} ⊂ H0 . 

We also assume from now on that q has zero mean, that is, 01 q dx = q0 = 0, since adding a constant to q shifts its entire spectrum, but does not affect its gap lengths. Using a contraction argument it is easy to show that the diagonal of Sn vanishes at a unique point αn near n2 π 2 . Lemma 5 For m  1 and n  m there exists a unique real analytic function

αn : B0m → C,

  αn − n2 π 2 

B0m




 such that λ − n2 π 2 − an (λ , ·) λ =αn ≡ 0 on B0m .

m2 , 4n

Given q ∈ H0 we replace its Fourier coefficients qn for |n| large enough by pn = cn (αn (q), q) = qn + . . . . The point is that

 Sn (αn , q) =

 0 −pn , −p−n 0

so these Fourier coefficients are well adapted to the lengths of the correponding gaps. More precisely, we define a map Pm : B0m → H0 by Pm (q) =

∑

|n| 0 we define a new function wε by (wε )n = min(wn , ε ε |n| ).

Spectral gaps of potentials in weighted Sobolev spaces

429

This is indeed a normalized, symmetric and submultiplicative function on Z , hence a weight. Moreover, if w is subexponential, then clearly (wε )n = wn ,

n , 1,

for any ε > 0. If we now choose first N sufficiently large, and then ε sufficiently small, we can arrange that TN pwε
TN pw
p0 , p − TN pwε
2p0 , for TN p = ∑|n|N pn e2n . Altogether, we have pwε
3p0
m/2. According to Proposition 7 we thus have q = Pm−1 (p) ∈ Hwε . But since wε has the same asymptotics as w we indeed have q = Pm−1 (p) ∈ Hwε = Hw , as we wanted to show. Essentially the same reasoning applies in the exponential case, with one important difference. If w is exponential, then (wε )n = ε ε |n| ,

n , 1.

We thus may only conclude that q = Pm−1 (p) ∈ Hwε = H0,ε  Hw . So we conclude that q is real analytic, but its width of analyticity may be smaller than what the weight w may suggest.

References 1. P. D JAKOV & B. M ITYAGIN , Smoothness of Schr¨odinger operator potential in the case of Gevrey type asymptotics of the gaps. J. Funct. Anal. 195 (2002), 89–128. 2. P. D JAKOV & B. M ITYAGIN , Spectral triangles of Schr¨odinger operators with complex potentials. Selecta Math. (N.S.) 9 (2003), 495–528. 3. P. D JAKOV & B. M ITYAGIN , Instability zones of one-dimensional periodic Schr¨odinger and Dirac operators. (Russian) Uspekhi Mat. Nauk 61 (2006), 77–182; translation in Russian Math. Surveys 61 (2006). 4. M. G. G ASYMOV, Spectral analysis of a class of second order nonselfadjoint differential operators. Funct. Anal. Appl. 14 (1980), 14–19. 5. H. H OCHSTADT, Estimates on the stability interval’s for the Hill’s equation. Proc. AMS 14 (1963), 930–932.

430

J. Pöschel

6. T. K APPELER & B. M ITYAGIN , Gap estimates of the spectrum of Hill’s equation and action variables for KdV. Trans. Amer. Math. Soc. 351 (1999), 619–646. 7. T. K APPELER & B. M ITYAGIN , Estimates for periodic and Dirichlet eigenvalues of the Schr¨odinger operator. SIAM J. Math. Anal. 33 (2001), 113–152. ˇ & I. O. O STROWSK˘I , A characterization of the spectrum of Hill’s operator. 8. V. A. M AR CENKO Math. USSR Sbornik 97 (1975), 493–554. 9. G. P ÓLYA & G. S ZEGo¨ , Problems and Theorems in Analysis I. Springer, New York, 1976. 10. J. Po¨ SCHEL , Hill’s potentials in weihtes Sobolev spaces and their spectral gaps. Preprint 2004, www.poschel.de/pbl.html. 11. J. J. S ANSUC & V. T KACHENKO , Spectral properties of non-selfadjoint Hill’s operators with smooth potentials. Algebraic and Geometric Methods in Mathematical Physics (Kaciveli, 1993), 371–385, Kluwer, 1996. 12. E. T RUBOWITZ , The inverse problem for periodic potentials. Comm. Pure Appl. Math. 30 (1977), 321–342.

On the well-posedness of the periodic KdV equation in high regularity classes Thomas Kappeler1 and Jürgen Pöschel2

Abstract We prove well-posedness results for the initial value problem of the periodic KdV equation in classes of high regularity solutions. More precisely, we consider the problem in weighted Sobolev spaces, which comprise classical Sobolev spaces, Gevrey spaces, and analytic spaces. We show that the initial value problem is well posed in all spaces with subexponential growth of Fourier coefficients, and ‘almost well posed’ in spaces with exponential growth of Fourier coefficients.

1 Results We consider the inital value problem for the periodic KdV equation,  u t=0 = u0 , ut = −uxxx + 6uux ,

(1)

where all functions are considered to be defined on T = R/Z . According to one of the first results in this direction due to Bona & Smith [5] this problem has a unique, global solution for any initial value in one of the standard Sobolev space Hm = H m (T, R) with m  2. That is, for each u0 ∈ Hm there exists a unique continuous curve

ϕ : R → Hm ,

t → ϕ (t, u0 )

solving the initial value in the sense defined below. Moreover, taken together they define a continous flow R × Hm → Hm ,

(t, u0 ) → ϕ (t, u0 ).

1 Institut

für Mathematik, Universität Zürich, Winterthurerstrasse 190, CH-8057 Zürich e-mail: [email protected]

2 Institut

für Analysis, Dynamik und Optimierung, Universität Stuttgart, Pfaffenwaldring 57 D-70569 Stuttgart e-mail: [email protected]

W. Craig (ed.), Hamiltonian Dynamical Systems and Applications, 431–441. c 2008 Springer Science + Business Media B.V.


431

432

T. Kappeler and J. Pöschel

Thus, the initial value problem is globally well-posed on Hm with m  2 in the sense of Hadamard: solutions exist for all time, are unique, and depend continuously on their initial values.

Well-Posedness Before we proceed we fix some notions. Let Hr = H r (T, R) be the usual Sobolev space of 1-periodic, real valued functions for real r  0. A continuous curve ϕ : I → Hr is called a solution of the initial value problem (1), if it solves (1) is the usual sense of distributions with ϕ (0) = u0 . It is called global, if I = R. We then say that the initial value problem (1) is globally well-posed in Hr , if it has a global solution for each initial value in Hr , and the resulting flow R × Hr → Hr ,

(t, u) → ϕ (t, u)

is continuous. Moreover, we call (1) globally uniformly well-posed in Hr , if it is globally well-posed, and for every compact interval I the map Hr → C0 (I, Hr ),

u → ϕ (·, u)

is uniformly continuous on bounded subsets of Hr with respect to the usual supnorm on the second space. Well-posedness in the spaces Hw introduced later is defined analogously.

Known Results Since the first results of Temam [31], Sjöberg [30] and Bona & Smith [5], the inital value problem for KdV and its well-posedness have been studied intensively. An excellent overview with a detailed bibliography is provided by the web site created by Colliander, Keel, Staffilani, Takaoka & Tao [11]. One focus has been on low regularity solutions in Sobolev spaces Hr with r
0. We mention the works [6–10, 20–22]. As a result, KdV is now known to be globally well-posed in Hr for every r  −1, and globally uniformly well-posed in Hr for every r  −1/2. Incidentally, it is an interesting phenomenon, that an equation can be globally well-posed, but not in a uniform way. In this paper we focus on high regularity solutions. These are solutions in a general class of weighted Sobolev spaces within H0 , that encompass analytic and Gevrey spaces, among others. Some results in this direction on the real line can be found in [4, 14]. But in general, the question of existence and well-posedness of solutions of nonlinear pdes of high regularity have not been widely considered. We this that this topic deserves to be studied in more depth, revealing important features of the nonlinear equation considered.

On the well-posedness of the periodic KdV equation in high regularity classes

433

Weighted Sobolev Spaces To state our results we first introduce weighted Sobolev spaces within the standard space H0 = L2 (T) of square-integrable functions u = ∑n∈Z un ε 2π inx . As in [16, 17] a weight is a normalized, symmetric and submultiplicative function w : Z → R. That is, for all integers n and m, we have wn  1,

w−n = wn ,

wn+m
wn wm .

We then define the weighted Sobolev spaces Hw := {u ∈ H0 : u2w := ∑n∈Z w2n |un |2 < ∞}. To give some examples, let n = 1 + |n|. The Sobolev weights nr , r  0, give rise to the usual Sobolev spaces Hr of 1-periodic, complex-valued functions. In particular, for nonnegative integers m we obtain the standard spaces Hm . The Abel weights1 nr ε a|n| with a > 0 define spaces Hr,a of functions in Hr , which are analytic on the complex strip |Im z| < a/2π and have traces in Hr on the boundary lines. The Gevrey weights σ

wn = nr ε a|n| ,

r  0, a > 0, 0 < σ < 1,

lie in between and give rise to the so called Gevrey spaces Hr,a,σ of smooth 1-periodic functions. Obviously, Hr,a = Hr,a,1  Hr,a,σ  Hr,a,0 = Hr . Since log wn is subadditive and nonnegative, the limit

χ (w) := lim

n→∞

log wn n

exists and is nonnegative [29, no. 98]. Naturally, we call a weight w exponential, if χ (w) > 0. We call w subexponential, if χ (w) = 0 with log wn /n converging to zero in an eventually monotone manner. This is not an exact dichotomy, but we are not aware of any interesting weight that does not belong to either class.

Theorems Theorem 1 The periodic KdV equation is globally uniformly well-posed in every space Hw with a subexponential weight w. That is, for each initial value u in one of 1

The term Abel weights is chosen to go along with Sobolev and Gevrey weights and has no deeper meaning.

434

T. Kappeler and J. Pöschel

these spaces Hw the associated Cauchy problem has a global solution t → ϕ t (u) in Hw , giving rise to a continuous flow (t, u) → ϕ t (u),

R × Hw → Hw ,

which is even uniformly continuous on bounded subsets of Hw . Indeed, the flow map is even analytic, see also [3]. For exponential weights the result is not as clear cut. Theorem 2 The periodic KdV equation is “almost” globally well-posed in every space Hw with an exponential weight w. That is, for each bounded subset B of Hw there exists 0 < ρ
1 such that the Cauchy problem for each initial value u ∈ B has ρ a global solution t → ϕ t (u) in Hw , giving rise to a continuous flow ρ

R × B → Hw ,

(t, u) → ϕ t (u). ρ

Here, wρ is the weight with (wρ )n = wn , which is again normalized, symmetric and submultiplicative. Thus, for initial values u in a bounded subset B of H0,a , say, (1) has a global solution in H0,ρ a with a fixed 0 < ρ
1. It is an open question, whether ρ can be chosen to be 1. For related results, see for example [1]. These results are not restricted to the standard KdV equation, but apply simultaneously to all equations in the KdV hierarchy, as defined for instance in [18]. The second KdV equation, for example, reads ut = uxxxxx − 10uuxxx − 20ux uxx + 30u2 ux . Such a hierarchy may be defined in a variety of ways, but this is immaterial here and does not affect the statement of the following theorem. Theorem 3 Theorems 1 and 2 also hold for every KdV equation in the KdV hierarchy, provided that in the case of Sobolev spaces Hr , r is sufficiently large. Our results naturally extend the KAM theory of Hamiltonian perturbations of KdV equations developed by Kuksin [23–25] and expounded in [18, 26]. Consider the perturbed KdV equation   ∂u d ∂H ∂K = +ε . ∂t dx ∂ u ∂u If K is real analytic in u with a gradient ∂ K/∂ u in some standard Sobolev space Hm , m  1, then KAM for KdV asserts the persistence of quasi-periodic solutions for sufficiently small ε = 0. Theorems 1 and 2 may now be extended as follows – for a more precise statement we refer to [19]. Theorem 4 Under sufficiently small Hamiltonian perturbations, the majority of the quasi-periodic solutions of the KdV equation persists, their regularity being only slightly less than the regularity of the perturbing term.

On the well-posedness of the periodic KdV equation in high regularity classes

435

These theorems are based on two observations. First, the periodic KdV equation is well known to be an infinite dimensional, integrable Hamiltonian system. As such, it even admits global Birkhoff coordinates (xn , yn )n1 defined as the cartesian counterpart to global action angle coordinates (In , θn )n1 . Second, there is a precise correspondence between the decay properties of the coordinates (xn , yn )n1 and the regularity properties of u. The link is provided by the spectral properties of the associated Hill operator d2 Lu = − 2 + u dx on the interval [0, 2] with periodic boundary conditions. In the rest of this note we describe this approach in more detail, but without lengthy proofs. These are given in [19].

2 Birkhoff Coordinates As is well known, the KdV equation can be written as an infinite dimensional Hamiltonian system ∂u d ∂H = ∂t dx ∂ u with Hamiltonian  1 H(u) = ( u2x + u3 ) dx. T 2 As a phase space one may take H0m = {u ∈ Hw : [u] :=



T u dx

= 0}

with m  1, as the KdV flow preserves mean values. The Poisson bracket proposed by Gardner,  ∂F d ∂G {F ,G} = dx, T ∂ u dx ∂ u then makes H0m into a nondegenerate Poisson manifold, such that ut = {u,H}. Next, we introduce the weighted sequence spaces hw = w × w with elements (x, y), where w = {x = (xn )n1 : x2w = ∑n1 w2n |xn |2 < ∞}. We endow hw with the standard Poisson structure, for which {xn ,ym } = δnm , while all other brackets vanish. To simplify notations, we further introduce √ w = {x ∈ w : ( nxn )n1 ∈ w }. hw = w × w ,

436

T. Kappeler and J. Pöschel

√ The extra weight n reflects the effect of the derivative d/dx in the Gardner bracket. The following theorem was first proven in [2, 3]. A quite different approach was first presented in [15], anda comprehensive exposition is given in [18]. Note that  H00 = u ∈ L2 (T) : [u] = 0 . Theorem 5 There exists a diffeomorphism

Ω : H00 → h0 with the following properties. (i) Ω is onto, bi-analytic, and takes the standard Poisson bracket into the Gardner bracket. (ii) The restriction of Ω to H0m , m  1, gives rise to a map Ω : H0m → hm  , which is again onto and bi-analytic. (iii) Ω introduces global Birkhoff coordinates for the KdV Hamiltonian on H01 . That is, on h1 the transformed KdV Hamiltonian H ◦ Ω −1 is a real analytic function of 1 In = (xn2 + y2n ), n  1. 2 (iv) The last statement also applies to every other Hamiltonian in the KdV hierarchy, if ‘1’ is replaced by ‘m’ with m sufficiently large. Denoting the transformed KdV Hamiltonian by the same symbol we thus obtain a real analytic Hamiltonian H = H(I1 , I2 , . . . ) on h1 . Its equations of motion are the classical ones, x˙n = Hyn ,

y˙n = −Hxn ,

n  1,

since the Poisson structure on h1 is the standard one. It is therefore evident that every solution of the KdV equation exists for all time, and is indeed almost periodic. More precisely, every solution winds around some underlying invariant torus TI = ∏ SIn ,

SIn = {xn2 + y2n = 2In },

n1

which is fixed by the actions of the initial position. The speed on the n-th circle SIn is determined by the n-th frequency

ωn = HIn (I1 , I2 , . . . ), and the entire flow is given by

ψ t (x, y) = (xn cos ωnt, yn sin ωnt)n1 . Obviously, ψ t preserves all weighted norms and thus all weighted spaces hw .

On the well-posedness of the periodic KdV equation in high regularity classes

437

To obtains our results about the well-posedness of the KdV equation, we now formulate two extensions of Theorem 5. First we consider subexponential weights. Theorem 6 For each subexponential weight w, the restriction of Ω to H0w gives rise to an onto, bi-analytic diffeomorphism Ω : H0w → hw . Proof (Proof of Theorem 1). Due to its symplectic nature, Ω maps solution curves t → ϕ t (u) in function space into solution curves t → ψ t (x, y) in sequence space with (x, y) = Ω (u). Since Ω is also a diffeomorphism between H0w and hw and ψ t preserves hw , the diagram Ω

u ∈ H0w ⏐ ⏐ ϕt